Normalize Trajectory Eval Score (#7668)

langchain/evaluation/__init__.py

@@ -37,7 +37,7 @@ name of the dataset to load.
 - Grading the accuracy of a response against ground truth answers: :class:`QAEvalChain <langchain.evaluation.qa.eval_chain.QAEvalChain>`
 - Comparing the output of two models: :class:`PairwiseStringEvalChain <langchain.evaluation.comparison.eval_chain.PairwiseStringEvalChain>` or :class:`LabeledPairwiseStringEvalChain <langchain.evaluation.comparison.eval_chain.LabeledPairwiseStringEvalChain>` when there is additionally a reference label.
 - Judging the efficacy of an agent's tool usage: :class:`TrajectoryEvalChain <langchain.evaluation.agents.trajectory_eval_chain.TrajectoryEvalChain>`
-- Checking whether an output complies with a set of criteria: :class:`CriteriaEvalChain <langchain.evaluation.criteria.eval_chain.CriteriaEvalChain>`
+- Checking whether an output complies with a set of criteria: :class:`CriteriaEvalChain <langchain.evaluation.criteria.eval_chain.CriteriaEvalChain>` or :class:`LabeledCriteriaEvalChain <langchain.evaluation.criteria.eval_chain.LabeledCriteriaEvalChain>` when there is additionally a reference label.
 - Computing semantic difference between a prediction and reference: :class:`EmbeddingDistanceEvalChain <langchain.evaluation.embedding_distance.base.EmbeddingDistanceEvalChain>` or between two predictions: :class:`PairwiseEmbeddingDistanceEvalChain <langchain.evaluation.embedding_distance.base.PairwiseEmbeddingDistanceEvalChain>` 
 - Measuring the string distance between a prediction and reference :class:`StringDistanceEvalChain <langchain.evaluation.string_distance.base.StringDistanceEvalChain>` or between two predictions :class:`PairwiseStringDistanceEvalChain <langchain.evaluation.string_distance.base.PairwiseStringDistanceEvalChain>`
 

langchain/evaluation/agents/trajectory_eval_chain.py

@@ -27,8 +27,12 @@ from langchain.tools.base import BaseTool
 
 
 class TrajectoryEval(NamedTuple):
-    score: int
+    """A named tuple containing the score and reasoning for a trajectory."""
+
+    score: float
+    """The score for the trajectory, normalized from 0 to 1.s"""
     reasoning: str
+    """The reasoning for the score."""
 
 
 class TrajectoryOutputParser(BaseOutputParser):
@@ -43,11 +47,11 @@ class TrajectoryOutputParser(BaseOutputParser):
             text (str): The output text to parse.
 
         Returns:
-            TrajectoryEval: A named tuple containing the score and reasoning.
+            TrajectoryEval: A named tuple containing the normalized score and reasoning.
 
         Raises:
             OutputParserException: If the score is not found in the output text or
-                if the score is not a digit in the range 1-5.
+                if the LLM's score is not a digit in the range 1-5.
         """
         if "Score:" not in text:
             raise OutputParserException(
@@ -66,8 +70,8 @@ class TrajectoryOutputParser(BaseOutputParser):
             raise OutputParserException(
                 f"Score is not a digit in the range 1-5: {text}"
             )
-
-        return TrajectoryEval(score=int(score_str), reasoning=reasoning)
+        normalized_score = (int(score_str) - 1) / 4
+        return TrajectoryEval(score=normalized_score, reasoning=reasoning)
 
 
 class TrajectoryEvalChain(AgentTrajectoryEvaluator, LLMEvalChain):
@@ -90,7 +94,7 @@ class TrajectoryEvalChain(AgentTrajectoryEvaluator, LLMEvalChain):
             \"\"\"Very helpful answers to geography questions.\"\"\"
             return f"{country}? IDK - We may never know {question}."
 
-        llm = ChatOpenAI(model="gpt-3.5-turbo-0613", temperature=0)
+        llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=0)
         agent = initialize_agent(
             tools=[geography_answers],
             llm=llm,

tests/unit_tests/evaluation/agents/test_eval_chain.py

@@ -70,7 +70,7 @@ def test_trajectory_eval_chain(
         agent_trajectory=intermediate_steps,
         prediction="I like pie.",
     )
-    assert res["score"] == 5
+    assert res["score"] == 1.0
     # Test when ref is provided
     res = chain.evaluate_agent_trajectory(
         input="What is your favorite food?",
@@ -78,7 +78,7 @@ def test_trajectory_eval_chain(
         prediction="I like pie.",
         reference="Paris",
     )
-    assert res["score"] == 1
+    assert res["score"] == 0.0
 
 
 def test_trajectory_eval_chain_no_tools(
@@ -97,14 +97,14 @@ def test_trajectory_eval_chain_no_tools(
         agent_trajectory=intermediate_steps,
         prediction="I like pie.",
     )
-    assert res["score"] == 5
+    assert res["score"] == 1.0
     res = chain.evaluate_agent_trajectory(
         input="What is your favorite food?",
         agent_trajectory=intermediate_steps,
         prediction="I like pie.",
         reference="Paris",
     )
-    assert res["score"] == 1
+    assert res["score"] == 0.0
 
 
 def test_old_api_works(intermediate_steps: List[Tuple[AgentAction, str]]) -> None:
@@ -123,7 +123,7 @@ def test_old_api_works(intermediate_steps: List[Tuple[AgentAction, str]]) -> Non
             "answer": "I like pie.",
         }
     )
-    assert res["score"] == 5
+    assert res["score"] == 1.0
 
     res = chain(
         {
@@ -133,4 +133,4 @@ def test_old_api_works(intermediate_steps: List[Tuple[AgentAction, str]]) -> Non
             "reference": "Paris",
         }
     )
-    assert res["score"] == 1
+    assert res["score"] == 0.0