You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
langchain/cookbook/causal_program_aided_langua...

693 lines
30 KiB
Plaintext

CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
{
"cells": [
{
"attachments": {},
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"cell_type": "markdown",
"id": "82f3f65d-fbcb-4e8e-b04b-959856283643",
"metadata": {},
"source": [
"# Causal program-aided language (CPAL) chain\n",
"\n",
"The CPAL chain builds on the recent PAL to stop LLM hallucination. The problem with the PAL approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination.\n",
"\n",
"The original [PR's description](https://github.com/langchain-ai/langchain/pull/6255) contains a full overview.\n",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"\n",
"Using the CPAL chain, the LLM translated this\n",
"\n",
" \"Tim buys the same number of pets as Cindy and Boris.\"\n",
" \"Cindy buys the same number of pets as Bill plus Bob.\"\n",
" \"Boris buys the same number of pets as Ben plus Beth.\"\n",
" \"Bill buys the same number of pets as Obama.\"\n",
" \"Bob buys the same number of pets as Obama.\"\n",
" \"Ben buys the same number of pets as Obama.\"\n",
" \"Beth buys the same number of pets as Obama.\"\n",
" \"If Obama buys one pet, how many pets total does everyone buy?\"\n",
"\n",
"\n",
"into this\n",
"\n",
"![complex-graph.png](/img/cpal_diagram.png).\n",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"\n",
"Outline of code examples demoed in this notebook.\n",
"\n",
"1. CPAL's value against hallucination: CPAL vs PAL \n",
" 1.1 Complex narrative \n",
" 1.2 Unanswerable math word problem \n",
"2. CPAL's three types of causal diagrams ([The Book of Why](https://en.wikipedia.org/wiki/The_Book_of_Why)). \n",
" 2.1 Mediator \n",
" 2.2 Collider \n",
" 2.3 Confounder "
]
},
{
"cell_type": "code",
"execution_count": 1,
"id": "1370e40f",
"metadata": {},
"outputs": [],
"source": [
"from IPython.display import SVG\n",
"from langchain_experimental.cpal.base import CPALChain\n",
"from langchain_experimental.pal_chain import PALChain\n",
"from langchain_openai import OpenAI\n",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"\n",
"llm = OpenAI(temperature=0, max_tokens=512)\n",
"cpal_chain = CPALChain.from_univariate_prompt(llm=llm, verbose=True)\n",
"pal_chain = PALChain.from_math_prompt(llm=llm, verbose=True)"
]
},
{
"cell_type": "markdown",
"id": "858a87d9-a9bd-4850-9687-9af4b0856b62",
"metadata": {},
"source": [
"## CPAL's value against hallucination: CPAL vs PAL\n",
"\n",
"Like PAL, CPAL intends to reduce large language model (LLM) hallucination.\n",
"\n",
"The CPAL chain is different from the PAL chain for a couple of reasons.\n",
"\n",
"CPAL adds a causal structure (or DAG) to link entity actions (or math expressions).\n",
"The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities.\n"
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
]
},
{
"cell_type": "markdown",
"id": "496403c5-d268-43ae-8852-2bd9903ce444",
"metadata": {},
"source": [
"### 1.1 Complex narrative\n",
"\n",
"Takeaway: PAL hallucinates, CPAL does not hallucinate."
]
},
{
"cell_type": "code",
"execution_count": 2,
"id": "d5dad768-2892-4825-8093-9b840f643a8a",
"metadata": {},
"outputs": [],
"source": [
"question = (\n",
" \"Tim buys the same number of pets as Cindy and Boris.\"\n",
" \"Cindy buys the same number of pets as Bill plus Bob.\"\n",
" \"Boris buys the same number of pets as Ben plus Beth.\"\n",
" \"Bill buys the same number of pets as Obama.\"\n",
" \"Bob buys the same number of pets as Obama.\"\n",
" \"Ben buys the same number of pets as Obama.\"\n",
" \"Beth buys the same number of pets as Obama.\"\n",
" \"If Obama buys one pet, how many pets total does everyone buy?\"\n",
")"
]
},
{
"cell_type": "code",
"execution_count": 3,
"id": "bbffa7a0-3c22-4a1d-ab2d-f230973073b0",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mdef solution():\n",
" \"\"\"Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?\"\"\"\n",
" obama_pets = 1\n",
" tim_pets = obama_pets\n",
" cindy_pets = obama_pets + obama_pets\n",
" boris_pets = obama_pets + obama_pets\n",
" total_pets = tim_pets + cindy_pets + boris_pets\n",
" result = total_pets\n",
" return result\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"'5'"
]
},
"execution_count": 3,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pal_chain.run(question)"
]
},
{
"cell_type": "code",
"execution_count": 4,
"id": "35a70d1d-86f8-4abc-b818-fbd083f072e9",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mstory outcome data\n",
" name code value depends_on\n",
"0 obama pass 1.0 []\n",
"1 bill bill.value = obama.value 1.0 [obama]\n",
"2 bob bob.value = obama.value 1.0 [obama]\n",
"3 ben ben.value = obama.value 1.0 [obama]\n",
"4 beth beth.value = obama.value 1.0 [obama]\n",
"5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob]\n",
"6 boris boris.value = ben.value + beth.value 2.0 [ben, beth]\n",
"7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris]\u001b[0m\n",
"\n",
"\u001b[36;1m\u001b[1;3mquery data\n",
"{\n",
" \"question\": \"how many pets total does everyone buy?\",\n",
" \"expression\": \"SELECT SUM(value) FROM df\",\n",
" \"llm_error_msg\": \"\"\n",
"}\u001b[0m\n",
"\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"13.0"
]
},
"execution_count": 4,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"cpal_chain.run(question)"
]
},
{
"cell_type": "code",
"execution_count": 5,
"id": "ccb6b2b0-9de6-4f66-a8fb-fc59229ee316",
"metadata": {},
"outputs": [
{
"data": {
"image/svg+xml": "<svg xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" width=\"292pt\" height=\"260pt\" viewBox=\"0.00 0.00 292.00 260.00\">\n<g id=\"graph0\" class=\"graph\" transform=\"scale(1 1) rotate(0) translate(4 256)\">\n<polygon fill=\"white\" stroke=\"transparent\" points=\"-4,4 -4,-256 288,-256 288,4 -4,4\"/>\n<!-- obama -->\n<g id=\"node1\" class=\"node\">\n<title>obama</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"137\" cy=\"-234\" rx=\"41.69\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"137\" y=\"-230.3\" font-family=\"Times,serif\" font-size=\"14.00\">obama</text>\n</g>\n<!-- bill -->\n<g id=\"node2\" class=\"node\">\n<title>bill</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"27\" cy=\"-162\" rx=\"27\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"27\" y=\"-158.3\" font-family=\"Times,serif\" font-size=\"14.00\">bill</text>\n</g>\n<!-- obama&#45;&gt;bill -->\n<g id=\"edge1\" class=\"edge\">\n<title>obama-&gt;bill</title>\n<path fill=\"none\" stroke=\"black\" d=\"M114.47,-218.67C97.08,-207.6 72.94,-192.23 54.42,-180.45\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"56.15,-177.4 45.84,-174.99 52.4,-183.31 56.15,-177.4\"/>\n</g>\n<!-- bob -->\n<g id=\"node3\" class=\"node\">\n<title>bob</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"100\" cy=\"-162\" rx=\"28\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"100\" y=\"-158.3\" font-family=\"Times,serif\" font-size=\"14.00\">bob</text>\n</g>\n<!-- obama&#45;&gt;bob -->\n<g id=\"edge2\" class=\"edge\">\n<title>obama-&gt;bob</title>\n<path fill=\"none\" stroke=\"black\" d=\"M128.04,-216.05C123.66,-207.77 118.3,-197.62 113.44,-188.42\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"116.39,-186.51 108.62,-179.31 110.2,-189.79 116.39,-186.51\"/>\n</g>\n<!-- ben -->\n<g id=\"node4\" class=\"node\">\n<title>ben</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"174\" cy=\"-162\" rx=\"28\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"174\" y=\"-158.3\" font-family=\"Times,serif\" font-size=\"14.00\">ben</text>\n</g>\n<!-- obama&#45;&gt;ben -->\n<g id=\"edge3\" class=\"edge\">\n<title>obama-&gt;ben</title>\n<path fill=\"none\" stroke=\"black\" d=\"M145.96,-216.05C150.34,-207.77 155.7,-197.62 160.56,-188.42\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"163.8,-189.79 165.38,-179.31 157.61,-186.51 163.8,-189.79\"/>\n</g>\n<!-- beth -->\n<g id=\"node5\" class=\"node\">\n<title>beth</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"252\" cy=\"-162\" rx=\"32\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"252\" y=\"-158.3\" font-family=\"Times,serif\" font-size=\"14.00\">beth</text>\n</g>\n<!-- obama&#45;&gt;beth -->\n<g id=\"edge4\" class=\"edge\">\n<title>obama-&gt;beth</title>\n<path fill=\"none\" stroke=\"black\" d=\"M160.27,-218.83C178.18,-207.94 203.04,-192.8 222.37,-181.04\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"224.36,-183.92 231.08,-175.73 220.72,-177.95 224.36,-183.92\"/>\n</g>\n<!-- cindy -->\n<g id=\"node6\" class=\"node\">\n<title>cindy</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"93\" cy=\"-90\" rx=\"36\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"93\" y=\"-86.3\" font-family=\"Times,serif\" font-size=\"14.00\">cindy</text>\n</g>\n<!-- bill&#45;&gt;cindy -->\n<g id=\"edge5\" class=\"edge\">\n<title>bill-&gt;cindy</title>\n<path fill=\"none\" stroke=\"black\" d=\"M41,-146.15C49.77,-136.85 61.25,-124.67 71.2,-114.12\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"73.79,-116.47 78.11,-106.8 68.7,-111.67 73.79,-116.47\"/>\n</g>\n<!-- bob&#45;&gt;cindy -->\n<g id=\"edge6\" class=\"edge\">\n<title>bob-&gt;cindy</title>\n<path fill=\"none\" stroke=\"black\" d=\"M98.27,-143.7C97.5,-135.98 96.57,-126.71 95.71,-118.11\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"99.19,-117.7 94.71,-108.1 92.22,-118.4 99.19,-117.7\"/>\n</g>\n<!-- boris -->\n<g id=\"node7\" class=\"node\">\n<title>boris</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"181\" cy=\"-90\" rx=\"34.5\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"181\" y=\"-86.3\" font-fa
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"text/plain": [
"<IPython.core.display.SVG object>"
]
},
"execution_count": 5,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# wait 20 secs to see display\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
"cpal_chain.draw(path=\"web.svg\")\n",
"SVG(\"web.svg\")"
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
]
},
{
"cell_type": "markdown",
"id": "1f6f345a-bb16-4e64-83c4-cbbc789a8325",
"metadata": {},
"source": [
"### Unanswerable math\n",
"\n",
"Takeaway: PAL hallucinates, where CPAL, rather than hallucinate, answers with _\"unanswerable, narrative question and plot are incoherent\"_"
]
},
{
"cell_type": "code",
"execution_count": 6,
"id": "068afd79-fd41-4ec2-b4d0-c64140dc413f",
"metadata": {},
"outputs": [],
"source": [
"question = (\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
" \"Jan has three times the number of pets as Marcia.\"\n",
" \"Marcia has two more pets than Cindy.\"\n",
" \"If Cindy has ten pets, how many pets does Barak have?\"\n",
")"
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
]
},
{
"cell_type": "code",
"execution_count": 7,
"id": "02f77db2-72e8-46c2-90b3-5e37ca42f80d",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mdef solution():\n",
" \"\"\"Jan has three times the number of pets as Marcia.Marcia has two more pets than Cindy.If Cindy has ten pets, how many pets does Barak have?\"\"\"\n",
" cindy_pets = 10\n",
" marcia_pets = cindy_pets + 2\n",
" jan_pets = marcia_pets * 3\n",
" result = jan_pets\n",
" return result\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"'36'"
]
},
"execution_count": 7,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pal_chain.run(question)"
]
},
{
"cell_type": "code",
"execution_count": 8,
"id": "925958de-e998-4ffa-8b2e-5a00ddae5026",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mstory outcome data\n",
" name code value depends_on\n",
"0 cindy pass 10.0 []\n",
"1 marcia marcia.value = cindy.value + 2 12.0 [cindy]\n",
"2 jan jan.value = marcia.value * 3 36.0 [marcia]\u001b[0m\n",
"\n",
"\u001b[36;1m\u001b[1;3mquery data\n",
"{\n",
" \"question\": \"how many pets does barak have?\",\n",
" \"expression\": \"SELECT name, value FROM df WHERE name = 'barak'\",\n",
" \"llm_error_msg\": \"\"\n",
"}\u001b[0m\n",
"\n",
"unanswerable, query and outcome are incoherent\n",
"\n",
"outcome:\n",
" name code value depends_on\n",
"0 cindy pass 10.0 []\n",
"1 marcia marcia.value = cindy.value + 2 12.0 [cindy]\n",
"2 jan jan.value = marcia.value * 3 36.0 [marcia]\n",
"query:\n",
"{'question': 'how many pets does barak have?', 'expression': \"SELECT name, value FROM df WHERE name = 'barak'\", 'llm_error_msg': ''}\n"
]
}
],
"source": [
"try:\n",
" cpal_chain.run(question)\n",
"except Exception as e_msg:\n",
" print(e_msg)"
]
},
{
"cell_type": "markdown",
"id": "095adc76",
"metadata": {},
"source": [
"### Basic math\n",
"\n",
"#### Causal mediator"
]
},
{
"cell_type": "code",
"execution_count": 9,
"id": "3ecf03fa-8350-4c4e-8080-84a307ba6ad4",
"metadata": {},
"outputs": [],
"source": [
"question = (\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
" \"Jan has three times the number of pets as Marcia. \"\n",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
" \"Marcia has two more pets than Cindy. \"\n",
" \"If Cindy has four pets, how many total pets do the three have?\"\n",
")"
]
},
{
"cell_type": "markdown",
"id": "74e49c47-3eed-4abe-98b7-8e97bcd15944",
"metadata": {},
"source": [
"---\n",
"PAL"
]
},
{
"cell_type": "code",
"execution_count": 10,
"id": "2e88395f-d014-4362-abb0-88f6800860bb",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mdef solution():\n",
" \"\"\"Jan has three times the number of pets as Marcia. Marcia has two more pets than Cindy. If Cindy has four pets, how many total pets do the three have?\"\"\"\n",
" cindy_pets = 4\n",
" marcia_pets = cindy_pets + 2\n",
" jan_pets = marcia_pets * 3\n",
" total_pets = cindy_pets + marcia_pets + jan_pets\n",
" result = total_pets\n",
" return result\u001b[0m\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"'28'"
]
},
"execution_count": 10,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"pal_chain.run(question)"
]
},
{
"cell_type": "markdown",
"id": "20ba6640-3d17-4b59-8101-aaba89d68cf4",
"metadata": {},
"source": [
"---\n",
"CPAL"
]
},
{
"cell_type": "code",
"execution_count": 11,
"id": "312a0943-a482-4ed0-a064-1e7a72e9479b",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mstory outcome data\n",
" name code value depends_on\n",
"0 cindy pass 4.0 []\n",
"1 marcia marcia.value = cindy.value + 2 6.0 [cindy]\n",
"2 jan jan.value = marcia.value * 3 18.0 [marcia]\u001b[0m\n",
"\n",
"\u001b[36;1m\u001b[1;3mquery data\n",
"{\n",
" \"question\": \"how many total pets do the three have?\",\n",
" \"expression\": \"SELECT SUM(value) FROM df\",\n",
" \"llm_error_msg\": \"\"\n",
"}\u001b[0m\n",
"\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"28.0"
]
},
"execution_count": 11,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"cpal_chain.run(question)"
]
},
{
"cell_type": "code",
"execution_count": 12,
"id": "4466b975-ae2b-4252-972b-b3182a089ade",
"metadata": {},
"outputs": [
{
"data": {
"image/svg+xml": "<svg xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" width=\"92pt\" height=\"188pt\" viewBox=\"0.00 0.00 92.49 188.00\">\n<g id=\"graph0\" class=\"graph\" transform=\"scale(1 1) rotate(0) translate(4 184)\">\n<polygon fill=\"white\" stroke=\"transparent\" points=\"-4,4 -4,-184 88.49,-184 88.49,4 -4,4\"/>\n<!-- cindy -->\n<g id=\"node1\" class=\"node\">\n<title>cindy</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"42.25\" cy=\"-162\" rx=\"36\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"42.25\" y=\"-158.3\" font-family=\"Times,serif\" font-size=\"14.00\">cindy</text>\n</g>\n<!-- marcia -->\n<g id=\"node2\" class=\"node\">\n<title>marcia</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"42.25\" cy=\"-90\" rx=\"42.49\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"42.25\" y=\"-86.3\" font-family=\"Times,serif\" font-size=\"14.00\">marcia</text>\n</g>\n<!-- cindy&#45;&gt;marcia -->\n<g id=\"edge1\" class=\"edge\">\n<title>cindy-&gt;marcia</title>\n<path fill=\"none\" stroke=\"black\" d=\"M42.25,-143.7C42.25,-135.98 42.25,-126.71 42.25,-118.11\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"45.75,-118.1 42.25,-108.1 38.75,-118.1 45.75,-118.1\"/>\n</g>\n<!-- jan -->\n<g id=\"node3\" class=\"node\">\n<title>jan</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"42.25\" cy=\"-18\" rx=\"27\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"42.25\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">jan</text>\n</g>\n<!-- marcia&#45;&gt;jan -->\n<g id=\"edge2\" class=\"edge\">\n<title>marcia-&gt;jan</title>\n<path fill=\"none\" stroke=\"black\" d=\"M42.25,-71.7C42.25,-63.98 42.25,-54.71 42.25,-46.11\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"45.75,-46.1 42.25,-36.1 38.75,-46.1 45.75,-46.1\"/>\n</g>\n</g>\n</svg>",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"text/plain": [
"<IPython.core.display.SVG object>"
]
},
"execution_count": 12,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# wait 20 secs to see display\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
"cpal_chain.draw(path=\"web.svg\")\n",
"SVG(\"web.svg\")"
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
]
},
{
"cell_type": "markdown",
"id": "29fa7b8a-75a3-4270-82a2-2c31939cd7e0",
"metadata": {},
"source": [
"### Causal collider"
]
},
{
"cell_type": "code",
"execution_count": 13,
"id": "618eddac-f0ef-4ab5-90ed-72e880fdeba3",
"metadata": {},
"outputs": [],
"source": [
"question = (\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
" \"Jan has the number of pets as Marcia plus the number of pets as Cindy. \"\n",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
" \"Marcia has no pets. \"\n",
" \"If Cindy has four pets, how many total pets do the three have?\"\n",
")"
]
},
{
"cell_type": "code",
"execution_count": 14,
"id": "a01563f3-7974-4de4-8bd9-0b7d710aa0d3",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mstory outcome data\n",
" name code value depends_on\n",
"0 marcia pass 0.0 []\n",
"1 cindy pass 4.0 []\n",
"2 jan jan.value = marcia.value + cindy.value 4.0 [marcia, cindy]\u001b[0m\n",
"\n",
"\u001b[36;1m\u001b[1;3mquery data\n",
"{\n",
" \"question\": \"how many total pets do the three have?\",\n",
" \"expression\": \"SELECT SUM(value) FROM df\",\n",
" \"llm_error_msg\": \"\"\n",
"}\u001b[0m\n",
"\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"8.0"
]
},
"execution_count": 14,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"cpal_chain.run(question)"
]
},
{
"cell_type": "code",
"execution_count": 15,
"id": "0fbe7243-0522-4946-b9a2-6e21e7c49a42",
"metadata": {},
"outputs": [
{
"data": {
"image/svg+xml": "<svg xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" width=\"182pt\" height=\"116pt\" viewBox=\"0.00 0.00 182.00 116.00\">\n<g id=\"graph0\" class=\"graph\" transform=\"scale(1 1) rotate(0) translate(4 112)\">\n<polygon fill=\"white\" stroke=\"transparent\" points=\"-4,4 -4,-112 178,-112 178,4 -4,4\"/>\n<!-- marcia -->\n<g id=\"node1\" class=\"node\">\n<title>marcia</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"42.25\" cy=\"-90\" rx=\"42.49\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"42.25\" y=\"-86.3\" font-family=\"Times,serif\" font-size=\"14.00\">marcia</text>\n</g>\n<!-- jan -->\n<g id=\"node2\" class=\"node\">\n<title>jan</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"90.25\" cy=\"-18\" rx=\"27\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"90.25\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">jan</text>\n</g>\n<!-- marcia&#45;&gt;jan -->\n<g id=\"edge1\" class=\"edge\">\n<title>marcia-&gt;jan</title>\n<path fill=\"none\" stroke=\"black\" d=\"M53.62,-72.41C59.57,-63.74 66.95,-52.97 73.53,-43.38\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"76.51,-45.21 79.28,-34.99 70.74,-41.26 76.51,-45.21\"/>\n</g>\n<!-- cindy -->\n<g id=\"node3\" class=\"node\">\n<title>cindy</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"138.25\" cy=\"-90\" rx=\"36\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"138.25\" y=\"-86.3\" font-family=\"Times,serif\" font-size=\"14.00\">cindy</text>\n</g>\n<!-- cindy&#45;&gt;jan -->\n<g id=\"edge2\" class=\"edge\">\n<title>cindy-&gt;jan</title>\n<path fill=\"none\" stroke=\"black\" d=\"M127.11,-72.77C121.09,-63.98 113.54,-52.96 106.83,-43.19\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"109.53,-40.94 100.99,-34.67 103.75,-44.89 109.53,-40.94\"/>\n</g>\n</g>\n</svg>",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"text/plain": [
"<IPython.core.display.SVG object>"
]
},
"execution_count": 15,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# wait 20 secs to see display\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
"cpal_chain.draw(path=\"web.svg\")\n",
"SVG(\"web.svg\")"
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
]
},
{
"cell_type": "markdown",
"id": "d4082538-ec03-44f0-aac3-07e03aad7555",
"metadata": {},
"source": [
"### Causal confounder"
]
},
{
"cell_type": "code",
"execution_count": 16,
"id": "83932c30-950b-435a-b328-7993ce8cc6bd",
"metadata": {},
"outputs": [],
"source": [
"question = (\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
" \"Jan has the number of pets as Marcia plus the number of pets as Cindy. \"\n",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
" \"Marcia has two more pets than Cindy. \"\n",
" \"If Cindy has four pets, how many total pets do the three have?\"\n",
")"
]
},
{
"cell_type": "code",
"execution_count": 17,
"id": "570de307-7c6b-4fdc-80c3-4361daa8a629",
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"\n",
"\n",
"\u001b[1m> Entering new chain...\u001b[0m\n",
"\u001b[32;1m\u001b[1;3mstory outcome data\n",
" name code value depends_on\n",
"0 cindy pass 4.0 []\n",
"1 marcia marcia.value = cindy.value + 2 6.0 [cindy]\n",
"2 jan jan.value = cindy.value + marcia.value 10.0 [cindy, marcia]\u001b[0m\n",
"\n",
"\u001b[36;1m\u001b[1;3mquery data\n",
"{\n",
" \"question\": \"how many total pets do the three have?\",\n",
" \"expression\": \"SELECT SUM(value) FROM df\",\n",
" \"llm_error_msg\": \"\"\n",
"}\u001b[0m\n",
"\n",
"\n",
"\u001b[1m> Finished chain.\u001b[0m\n"
]
},
{
"data": {
"text/plain": [
"20.0"
]
},
"execution_count": 17,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"cpal_chain.run(question)"
]
},
{
"cell_type": "code",
"execution_count": 18,
"id": "00375615-6b6d-4357-bdb8-f64f682f7605",
"metadata": {},
"outputs": [
{
"data": {
"image/svg+xml": "<svg xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" width=\"121pt\" height=\"188pt\" viewBox=\"0.00 0.00 120.99 188.00\">\n<g id=\"graph0\" class=\"graph\" transform=\"scale(1 1) rotate(0) translate(4 184)\">\n<polygon fill=\"white\" stroke=\"transparent\" points=\"-4,4 -4,-184 116.99,-184 116.99,4 -4,4\"/>\n<!-- cindy -->\n<g id=\"node1\" class=\"node\">\n<title>cindy</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"77.25\" cy=\"-162\" rx=\"36\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"77.25\" y=\"-158.3\" font-family=\"Times,serif\" font-size=\"14.00\">cindy</text>\n</g>\n<!-- marcia -->\n<g id=\"node2\" class=\"node\">\n<title>marcia</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"42.25\" cy=\"-90\" rx=\"42.49\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"42.25\" y=\"-86.3\" font-family=\"Times,serif\" font-size=\"14.00\">marcia</text>\n</g>\n<!-- cindy&#45;&gt;marcia -->\n<g id=\"edge1\" class=\"edge\">\n<title>cindy-&gt;marcia</title>\n<path fill=\"none\" stroke=\"black\" d=\"M68.95,-144.41C64.87,-136.25 59.86,-126.22 55.28,-117.07\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"58.33,-115.34 50.72,-107.96 52.07,-118.47 58.33,-115.34\"/>\n</g>\n<!-- jan -->\n<g id=\"node3\" class=\"node\">\n<title>jan</title>\n<ellipse fill=\"none\" stroke=\"black\" cx=\"77.25\" cy=\"-18\" rx=\"27\" ry=\"18\"/>\n<text text-anchor=\"middle\" x=\"77.25\" y=\"-14.3\" font-family=\"Times,serif\" font-size=\"14.00\">jan</text>\n</g>\n<!-- cindy&#45;&gt;jan -->\n<g id=\"edge2\" class=\"edge\">\n<title>cindy-&gt;jan</title>\n<path fill=\"none\" stroke=\"black\" d=\"M83.73,-144.1C87.32,-133.84 91.42,-120.36 93.25,-108 95.58,-92.17 95.58,-87.83 93.25,-72 91.95,-63.21 89.5,-53.86 86.91,-45.5\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"90.19,-44.29 83.73,-35.9 83.55,-46.49 90.19,-44.29\"/>\n</g>\n<!-- marcia&#45;&gt;jan -->\n<g id=\"edge3\" class=\"edge\">\n<title>marcia-&gt;jan</title>\n<path fill=\"none\" stroke=\"black\" d=\"M50.72,-72.06C54.86,-63.77 59.94,-53.62 64.53,-44.42\"/>\n<polygon fill=\"black\" stroke=\"black\" points=\"67.75,-45.82 69.09,-35.31 61.49,-42.69 67.75,-45.82\"/>\n</g>\n</g>\n</svg>",
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
"text/plain": [
"<IPython.core.display.SVG object>"
]
},
"execution_count": 18,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"# wait 20 secs to see display\n",
Fix `make docs_build` and related scripts (#7276) **Description: a description of the change** Fixed `make docs_build` and related scripts which caused errors. There are several changes. First, I made the build of the documentation and the API Reference into two separate commands. This is because it takes less time to build. The commands for documents are `make docs_build`, `make docs_clean`, and `make docs_linkcheck`. The commands for API Reference are `make api_docs_build`, `api_docs_clean`, and `api_docs_linkcheck`. It looked like `docs/.local_build.sh` could be used to build the documentation, so I used that. Since `.local_build.sh` was also building API Rerefence internally, I removed that process. `.local_build.sh` also added some Bash options to stop in error or so. Futher more added `cd "${SCRIPT_DIR}"` at the beginning so that the script will work no matter which directory it is executed in. `docs/api_reference/api_reference.rst` is removed, because which is generated by `docs/api_reference/create_api_rst.py`, and added it to .gitignore. Finally, the description of CONTRIBUTING.md was modified. **Issue: the issue # it fixes (if applicable)** https://github.com/hwchase17/langchain/issues/6413 **Dependencies: any dependencies required for this change** `nbdoc` was missing in group docs so it was added. I installed it with the `poetry add --group docs nbdoc` command. I am concerned if any modifications are needed to poetry.lock. I would greatly appreciate it if you could pay close attention to this file during the review. **Tag maintainer** - General / Misc / if you don't know who to tag: @baskaryan If this PR needs any additional changes, I'll be happy to make them! --------- Co-authored-by: Bagatur <baskaryan@gmail.com>
11 months ago
"cpal_chain.draw(path=\"web.svg\")\n",
"SVG(\"web.svg\")"
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
]
},
{
"cell_type": "code",
"execution_count": 19,
"id": "255683de-0c1c-4131-b277-99d09f5ac1fc",
"metadata": {},
"outputs": [],
"source": [
"%load_ext autoreload\n",
"%autoreload 2"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "Python 3 (ipykernel)",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.11.3"
CPAL (#6255) # Causal program-aided language (CPAL) chain ## Motivation This builds on the recent [PAL](https://arxiv.org/abs/2211.10435) to stop LLM hallucination. The problem with the [PAL](https://arxiv.org/abs/2211.10435) approach is that it hallucinates on a math problem with a nested chain of dependence. The innovation here is that this new CPAL approach includes causal structure to fix hallucination. For example, using the below word problem, PAL answers with 5, and CPAL answers with 13. "Tim buys the same number of pets as Cindy and Boris." "Cindy buys the same number of pets as Bill plus Bob." "Boris buys the same number of pets as Ben plus Beth." "Bill buys the same number of pets as Obama." "Bob buys the same number of pets as Obama." "Ben buys the same number of pets as Obama." "Beth buys the same number of pets as Obama." "If Obama buys one pet, how many pets total does everyone buy?" The CPAL chain represents the causal structure of the above narrative as a causal graph or DAG, which it can also plot, as shown below. ![complex-graph](https://github.com/hwchase17/langchain/assets/367522/d938db15-f941-493d-8605-536ad530f576) . The two major sections below are: 1. Technical overview 2. Future application Also see [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 1. Technical overview ### CPAL versus PAL Like [PAL](https://arxiv.org/abs/2211.10435), CPAL intends to reduce large language model (LLM) hallucination. The CPAL chain is different from the PAL chain for a couple of reasons. * CPAL adds a causal structure (or DAG) to link entity actions (or math expressions). * The CPAL math expressions are modeling a chain of cause and effect relations, which can be intervened upon, whereas for the PAL chain math expressions are projected math identities. PAL's generated python code is wrong. It hallucinates when complexity increases. ```python def solution(): """Tim buys the same number of pets as Cindy and Boris.Cindy buys the same number of pets as Bill plus Bob.Boris buys the same number of pets as Ben plus Beth.Bill buys the same number of pets as Obama.Bob buys the same number of pets as Obama.Ben buys the same number of pets as Obama.Beth buys the same number of pets as Obama.If Obama buys one pet, how many pets total does everyone buy?""" obama_pets = 1 tim_pets = obama_pets cindy_pets = obama_pets + obama_pets boris_pets = obama_pets + obama_pets total_pets = tim_pets + cindy_pets + boris_pets result = total_pets return result # math result is 5 ``` CPAL's generated python code is correct. ```python story outcome data name code value depends_on 0 obama pass 1.0 [] 1 bill bill.value = obama.value 1.0 [obama] 2 bob bob.value = obama.value 1.0 [obama] 3 ben ben.value = obama.value 1.0 [obama] 4 beth beth.value = obama.value 1.0 [obama] 5 cindy cindy.value = bill.value + bob.value 2.0 [bill, bob] 6 boris boris.value = ben.value + beth.value 2.0 [ben, beth] 7 tim tim.value = cindy.value + boris.value 4.0 [cindy, boris] query data { "question": "how many pets total does everyone buy?", "expression": "SELECT SUM(value) FROM df", "llm_error_msg": "" } # query result is 13 ``` Based on the comments below, CPAL's intended location in the library is `experimental/chains/cpal` and PAL's location is`chains/pal`. ### CPAL vs Graph QA Both the CPAL chain and the Graph QA chain extract entity-action-entity relations into a DAG. The CPAL chain is different from the Graph QA chain for a few reasons. * Graph QA does not connect entities to math expressions * Graph QA does not associate actions in a sequence of dependence. * Graph QA does not decompose the narrative into these three parts: 1. Story plot or causal model 4. Hypothetical question 5. Hypothetical condition ### Evaluation Preliminary evaluation on simple math word problems shows that this CPAL chain generates less hallucination than the PAL chain on answering questions about a causal narrative. Two examples are in [this jupyter notebook](https://github.com/borisdev/langchain/blob/master/docs/extras/modules/chains/additional/cpal.ipynb) doc. ## 2. Future application ### "Describe as Narrative, Test as Code" The thesis here is that the Describe as Narrative, Test as Code approach allows you to represent a causal mental model both as code and as a narrative, giving you the best of both worlds. #### Why describe a causal mental mode as a narrative? The narrative form is quick. At a consensus building meeting, people use narratives to persuade others of their causal mental model, aka. plan. You can share, version control and index a narrative. #### Why test a causal mental model as a code? Code is testable, complex narratives are not. Though fast, narratives are problematic as their complexity increases. The problem is LLMs and humans are prone to hallucination when predicting the outcomes of a narrative. The cost of building a consensus around the validity of a narrative outcome grows as its narrative complexity increases. Code does not require tribal knowledge or social power to validate. Code is composable, complex narratives are not. The answer of one CPAL chain can be the hypothetical conditions of another CPAL Chain. For stochastic simulations, a composable plan can be integrated with the [DoWhy library](https://github.com/py-why/dowhy). Lastly, for the futuristic folk, a composable plan as code allows ordinary community folk to design a plan that can be integrated with a blockchain for funding. An explanation of a dependency planning application is [here.](https://github.com/borisdev/cpal-llm-chain-demo) --- Twitter handle: @boris_dev --------- Co-authored-by: Boris Dev <borisdev@Boriss-MacBook-Air.local>
11 months ago
}
},
"nbformat": 4,
"nbformat_minor": 5
}