reflexion-human-eval/reflexion.py
2023-04-03 16:57:54 -04:00

87 lines
2.9 KiB
Python

from utils import enumerate_resume, write_jsonl
from executors import executor_factory
from generators import generator_factory
from typing import List
def run_reflexion(
dataset: List[dict],
model: str,
language: str,
max_iters: int,
pass_at_k: int,
log_path: str,
verbose: bool
) -> None:
exe = executor_factory(language)
gen = generator_factory(language)
num_items = len(dataset)
num_success = 0
for i, item in enumerate_resume(dataset, log_path):
cur_pass = 0
is_solved = False
reflections = []
cur_func_impl = ""
while cur_pass < pass_at_k and not is_solved:
tests_i = gen.internal_tests(item["prompt"], model, 1)
# first attempt
cur_func_impl = gen.func_impl(item["prompt"], model, "simple")
assert isinstance(cur_func_impl, str)
is_passing, feedback, _ = exe.execute(cur_func_impl, tests_i)
# if solved, exit early
if is_passing:
is_passing = exe.evaluate(
item["entry_point"], cur_func_impl, item["test"], timeout=10)
is_solved = is_passing
num_success += int(is_passing)
break
# use self-reflection to iteratively improve
cur_iter = 1
cur_feedback = feedback
while cur_iter < max_iters:
# get self-reflection
reflection = gen.self_reflection(
cur_func_impl, cur_feedback, model)
reflections += [reflection]
# apply self-reflection in the next attempt
cur_func_impl = gen.func_impl(
func_sig=item["prompt"],
model=model,
strategy="reflexion",
prev_func_impl=cur_func_impl,
feedback=cur_feedback,
self_reflection=reflection
)
assert isinstance(cur_func_impl, str)
# check if all internal unit tests pass
is_passing, cur_feedback, _ = exe.execute(cur_func_impl, tests_i)
# if solved, check if it passes the real tests, exit early
if is_passing or cur_iter == max_iters - 1:
is_passing = exe.evaluate(
item["entry_point"], cur_func_impl, item["test"], timeout=10)
if is_passing:
item["solution"] = cur_func_impl
is_solved = True
num_success += 1
break
cur_iter += 1
cur_pass += 1
item["is_solved"] = is_solved
item["reflections"] = reflections
item["solution"] = cur_func_impl
write_jsonl(log_path, [item], append=True)
if verbose:
print(
f'completed {i+1}/{num_items}: acc = {round(num_success/(i+1), 2)}')