One Token to Fool LLM-as-a-Judge
Paper • 2507.08794 • Published • 32
Datasets:
messages listlengths 3 3 | verdict stringclasses 2
values | score int64 1 5 |
|---|---|---|
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 4 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 3 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 4 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 3 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 4 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 4 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 1 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | PASS | 5 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 4 |
[
{
"role": "system",
"content": "You are a strict, careful verifier. You are given a TASK, a scoring CRITERION, an optional REFERENCE answer, and a candidate RESPONSE. Decide whether the RESPONSE satisfies the CRITERION.\nFirst write a brief critique citing concrete evidence from the RESPONSE. Then output th... | FAIL | 2 |
End of preview. Expand in Data Studio
verifier-debias-v2 — de-biased SFT data for a generative verifier
SFT data to train a generative verifier (GenRM, arXiv:2408.15240)
from Qwen/Qwen2.5-7B-Instruct. Each row is a conversational example
(messages = system + user + gold assistant) plus a verdict (PASS/FAIL) and
the underlying 1–5 score. The assistant target is a critique ending in
Verdict: PASS / Verdict: FAIL.
Built from prometheus-eval/Feedback-Collection with three literature-backed de-biasing fixes over v1:
- Stricter PASS threshold —
score == 5 → PASS(v1 usedscore ≥ 4), to stop the verifier rubber-stamping borderline (score 2–3) responses. - Fluent-but-wrong hard negatives — confident, well-written but incorrect responses labelled FAIL (OffsetBias 2407.06551, Master-RM 2507.08794).
- Length-balancing — decouple verdict from response length (RISE-Judge 2502.11689).
The held-out test split (2000 rows: 536 PASS / 1464 FAIL) is the hard
evaluation distribution.
Why this matters: the key finding
Use case is a cheap open-weights verifier for agent outputs (à la LangChain × Harvey). The priority metric is false-pass rate (approving a bad response — the costly error).
| Model | Test split | Agreement | False-pass | False-fail |
|---|---|---|---|---|
| v1 (qwen2.5-7b-verifier) | v1-test (score≥4→PASS) | 83.5% | 29.8% | 3.3% |
| v2 (qwen2.5-7b-verifier-v2) | v1-test | 80.4% | 23.5% | 15.7% |
| v2 | v2-test (this dataset, hard) | 69.3% | 39.5% | 6.7% |
De-biasing SFT lowered false-pass on the matched distribution (29.8%→23.5%) but could not fix the hard fluent-but-wrong split (39.5%). A full inference-time threshold sweep on P(PASS) confirmed this is not a calibration problem: the minimum achievable false-pass on v2-test across all thresholds is 39.1% — the model is confidently fooled. Per RISE-Judge, SFT plateaus on this "Chat-Hard" failure mode and a contrastive DPO stage is required. That is implemented here (see scripts) as v3.
Scripts in this repo
| Script | Purpose |
|---|---|
eval_v2_both_splits.py |
Load one model, eval on BOTH test splits (greedy verdict). |
eval_pointwise.py / eval_own_split.py |
Single model × single split eval. |
eval_threshold_sweep.py |
Continuous P(PASS) scoring + threshold sweep; finds the operating point for a target false-pass. |
mine_dpo_pairs.py |
RISE-Judge pair mining: sample critiques from the SFT model, keep fooled-PASS as rejected vs gold FAIL as chosen (+ recall-anchor PASS pairs) → verifier-dpo-v3. |
train_dpo_v3.py |
DPO (RPO variant, rpo_alpha) on top of the v2 SFT checkpoint → qwen2.5-7b-verifier-v3. |
All eval scripts pin vllm==0.6.6.post1, transformers==4.47.1, tokenizer from
base Qwen, temp 0.0.
Usage
from datasets import load_dataset
ds = load_dataset("narcolepticchicken/verifier-debias-v2")
References
GenRM 2408.15240 · Prometheus 2 2405.01535 · OffsetBias 2407.06551 · Master-RM 2507.08794 · RISE-Judge 2502.11689
Generated by ML Intern
Generated by ML Intern.
- Downloads last month
- 98