tdprm-results / code /v3 /tdprm /scripts /smoke_v3_reader.py
JoshuaW1997's picture
Super-squash branch 'main' using huggingface_hub
575b0ec
Raw History Blame Contribute Delete
3.2 kB
"""Reader smoke test (TDPRM_V3_EXPERIMENTS.md, section 3.2) against a running belief server.
python -m tdprm.scripts.smoke_v3_reader --url http://127.0.0.1:$PORT
Prints the three belief rows and the acceptance checks.
"""
import argparse
from ..credit import v3
from ..rl.provider_v3 import BeliefClient
MATH_MSGS = [{"role": "system", "content": "Solve the math problem step by step. Start each reasoning step with 'Step 1:', "
"'Step 2:', etc. Finish with a separate line 'Final answer: \\boxed{<answer>}'."},
{"role": "user", "content": "A shop sells pens at 3 dollars each and notebooks at 5 dollars each. Mia buys 4 "
"pens and 3 notebooks. How many dollars does she spend?"}]
MATH_STEPS = ["Step 1: The pens cost 4 x 3 = 12 dollars.", "Step 2: The notebooks cost 3 x 5 = 15 dollars.",
"Step 3: In total she spends 12 + 15 = 27 dollars."]
PQ_MSGS = [{"role": "system", "content": "Answer the logical reasoning question using the given context. Solve "
"step-by-step. Start each reasoning step with 'Step 1:', 'Step 2:', etc. Finish with a separate line "
"'Final answer: <option letter and text>'."},
{"role": "user", "content": "Context: Every member of the chess club practices on Sundays. Anyone who "
"practices on Sundays misses the Sunday market. Lena is a member of the chess club.\n\nQuestion: Based on "
"the context, is the following statement true, false, or uncertain? Lena misses the Sunday market.\n\n"
"Options:\nA) True\nB) False\nC) Uncertain"}]
PQ_STEPS = ["Step 1: Lena is a member of the chess club, so she practices on Sundays.",
"Step 2: Anyone who practices on Sundays misses the Sunday market, so Lena misses it. The statement is True."]
def main():
ap = argparse.ArgumentParser()
ap.add_argument("--url", required=True)
args = ap.parse_args()
c = BeliefClient(args.url, 120)
m_pre, _ = v3.build_prefixes(MATH_STEPS, [1, 2, 3], "off")
l_pre, k = v3.build_prefixes(PQ_STEPS, [1, 2], "labels", "A")
out = c.belief([{"messages": MATH_MSGS, "prefixes": m_pre, "answer": "27", "mode": "text", "labels": ["A", "B", "C"]},
{"messages": PQ_MSGS, "prefixes": l_pre, "answer": "A", "mode": "label", "labels": ["A", "B", "C"]},
{"messages": PQ_MSGS, "prefixes": l_pre, "answer": "B", "mode": "label", "labels": ["A", "B", "C"]}])
print("text (gold 27): Phi stub, s1, s2, s3 =", [round(x, 3) for x in out[0]])
print("label (gold A) : Phi stub, s1, s2 =", [round(x, 3) for x in out[1]], "| masked:", k)
print("label (gold B) : Phi stub, s1, s2 =", [round(x, 3) for x in out[2]])
checks = {"every value in [0, 1]": all(0.0 <= x <= 1.0 for row in out for x in row),
"masked = 1": k == 1,
"gold-B ~0 where gold-A ~1": all(b < 0.1 for a, b in zip(out[1], out[2]) if a > 0.9),
"text row ends near 1": out[0][-1] > 0.9}
for name, ok in checks.items():
print(f" {'PASS' if ok else 'FAIL'} {name}")
print("smoke", "PASS" if all(checks.values()) else "FAIL")
if __name__ == "__main__":
main()