claude
c098933f25
Sync course wiki / sync-wiki (push) Successful in 4s
Co-authored-by: claude <claude@jpaul.io> Co-committed-by: claude <claude@jpaul.io>
80 lines
3.4 KiB
Python
80 lines
3.4 KiB
Python
"""LLM-as-judge: the pattern, and its limits, in one file.
|
|
|
|
Some agent output can't be graded by `==`. "Is this commit message clear?" or
|
|
"Does this PR description actually explain the change?" has no exact answer. The
|
|
common move is to ask *another* model to grade it. This file shows the shape of
|
|
that grader and is deliberately honest about what it can't do.
|
|
|
|
It is vendor-agnostic by design. Point it at whatever model endpoint you already
|
|
use by setting two environment variables; if they're not set, it abstains rather
|
|
than pretending. NOTHING here pins a provider.
|
|
|
|
EVAL_JUDGE_URL # an OpenAI-style /chat/completions-compatible endpoint, or your own
|
|
EVAL_JUDGE_KEY # the bearer token for it
|
|
EVAL_JUDGE_MODEL # the model name to ask for
|
|
|
|
Run it standalone to grade one sample:
|
|
python llm_judge.py "Add count command" "fix"
|
|
"""
|
|
|
|
import json
|
|
import os
|
|
import sys
|
|
import urllib.request
|
|
|
|
RUBRIC = """You are grading one piece of agent output against a rubric.
|
|
Score 1 if the commit message clearly and specifically describes the change.
|
|
Score 0 if it is vague, generic, or could describe almost any change.
|
|
Reply with ONLY a JSON object: {"score": 0 or 1, "reason": "<one short sentence>"}
|
|
"""
|
|
|
|
|
|
def judge(candidate_text: str) -> dict:
|
|
url = os.environ.get("EVAL_JUDGE_URL")
|
|
key = os.environ.get("EVAL_JUDGE_KEY")
|
|
model = os.environ.get("EVAL_JUDGE_MODEL")
|
|
if not (url and key and model):
|
|
return {"score": None, "reason": "judge not configured; abstaining (set EVAL_JUDGE_* to enable)"}
|
|
|
|
payload = json.dumps({
|
|
"model": model,
|
|
"temperature": 0, # determinism matters for a grader; you want repeatable scores
|
|
"messages": [
|
|
{"role": "system", "content": RUBRIC},
|
|
{"role": "user", "content": f"Output to grade:\n{candidate_text}"},
|
|
],
|
|
}).encode()
|
|
|
|
req = urllib.request.Request(
|
|
url,
|
|
data=payload,
|
|
headers={"Authorization": f"Bearer {key}", "Content-Type": "application/json"},
|
|
)
|
|
with urllib.request.urlopen(req) as resp:
|
|
body = json.loads(resp.read())
|
|
content = body["choices"][0]["message"]["content"]
|
|
return json.loads(content)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
sample = sys.argv[1] if len(sys.argv) > 1 else "fix stuff"
|
|
print(judge(sample))
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# READ THIS BEFORE TRUSTING A SCORE FROM HERE.
|
|
#
|
|
# An LLM judge is a model grading a model. Its failure modes are real:
|
|
# - Correlated blind spots: the judge can share the candidate's confusion, so
|
|
# a wrong answer gets a passing grade because both models are wrong the same way.
|
|
# - Bias: judges favor longer, more confident, or first-presented answers
|
|
# regardless of correctness. Hold position and length constant when you can.
|
|
# - Drift: change the judge model and your scores move even though nothing
|
|
# about the candidate changed. The ruler is itself made of rubber.
|
|
#
|
|
# So: use a programmatic grader (run_eval.py) wherever a deterministic check is
|
|
# possible; that is most of the time. Reach for an LLM judge only for genuinely
|
|
# open-ended output, and CALIBRATE it first: hand-label ~20 examples yourself,
|
|
# run the judge on them, and confirm it agrees with you before you let it gate
|
|
# anything. An uncalibrated judge is a vibe with a number attached.
|
|
# ---------------------------------------------------------------------------
|