xiaoyuanwang/ifbench-audit

IFBench Checker Audit

Labeled data from an audit of the 58 IFBench instruction checkers that Marin runs through its Evalchemy fork (marin#10060). It has the audit's fixtures, findings, and interpretation decisions, two rounds of blinded review with both reviewers' labels, and the verdicts of six graders on a model panel of 68,400 responses.

The code, the report, and the scripts behind every number are in XiaoyuanWang111/ifbench-audit at a31c684. scripts/build_dataset.py builds this dataset from that commit. It also recomputes both rounds' error rates and inter-reviewer kappa from the written rows and checks them against the audit's scores.

Configs

Config	Rows	One row is
fixtures	143	A response labeled pass or fail for one instruction, from the instruction text alone
findings	50	A checker defect (F1–F43) or a protocol problem (G1–G7)
decisions	18	A question that the instruction text leaves open: 15 with a proposed reading, 3 without
blind_review	764	One blinded case: a response, one instruction, the graders' verdicts, and two reviewers' labels
blind_review_strata	6	The population and sample size of one stratum of one review round
panel_verdicts	68,400	One panel response, with each constraint's verdict under six graders

Graders

• marin_strict and marin_loose: Marin's checkers in Evalchemy's eval/chat_benchmarks/IFBench, whose checker code is unchanged from ab7ae27 to 09d27c9. The audit's reimplementation of Evalchemy's strict and loose grading matched Evalchemy on all 5,848 verdicts it was compared on: the greedy run of every configuration and upstream's sample responses.
• upstream_strict and upstream_loose: upstream IFBench at 1c40f0c.
• fixes_only: the audit's validated checkers with the official reading of every decision. They differ from Marin strict only where a defect is fixed.
• proposed: the validated checkers with the proposed reading of every decision.

Loose grading accepts a constraint if any of eight variants of the response passes: the response itself, the response with * removed, and each of those with the first line, the last line, or both dropped.

Fields

fixtures: id (task family plus a content hash), family, name, instruction_id, kwargs_json (the checker arguments as a JSON string), response, expected (the contract label), finding and decision (the finding or decision the fixture demonstrates), note, content_sha256.

findings: id, kind (checker or protocol), instruction_ids, n_test_constraints (constraints of these types in IFBench_test), defect, counterexample, decisions (decision entries named in the row), rests_on_decision (true for the nine findings that depend on a decision), marin_status and upstream_status (defective or fixed).

decisions: id, question, official_reading, proposal. The three open questions have neither of the last two.

blind_review: case_id, round, source (panel or fixture), stratum (disagree, agree, or fixture), config and prompt_key (panel cases), fixture_name, fixture_expected, finding, and decision (fixture cases), instruction_id, instruction (the text the reviewers judged), prompt, response (full text), marin_strict, fixes_only, and proposed (verdicts when the round was frozen; empty for fixtures), then reviewer_a_* and reviewer_b_* (verdict, lean, confidence, reason), and consensus (the shared lean, or empty when the reviewers' leans differ).

blind_review_strata: round, stratum, population (the number of panel verdicts the stratum stands for), sampled.

panel_verdicts: config, run, prompt_key (the IFBench_test key), response_sha256, extraction (ok or reasoning_unclosed), tokens, finish_reason, truncated, repeat_share (the share of repeated word 4-grams), degenerate (repeat_share of at least 0.5), infrastructure, and constraints: per constraint, instruction_id, the six graders' verdicts, proposed_reason, and crash (whether Marin's or upstream's checker raised an exception; none did on the panel). The responses are in release v0.1 of the repository. response_sha256 is the sha256 of each generation record's response field, the graded answer; raw_response keeps any reasoning.

How the data was made

The panel has 16 configurations of 14 open models (7B to 32B): Qwen3-8B and Qwen3-14B run with thinking on and off. Responses were generated with vLLM 0.30.0 on one NVIDIA H100 80GB GPU from allenai/IFBench_test at 2e8a48de, with each model's chat template and at most 1,024 new tokens (8,192 with thinking). run names the decoding run:

• greedy, greedy_repeat (the same prompt order again), and greedy_shuffled (shuffled prompt order);
• bi_greedy and bi_greedy_shuffled: greedy with vLLM's batch-invariant kernels (VLLM_BATCH_INVARIANT=1);
• sample0 to sample11: temperature 0.7 with seed 10060 + i, eight samples for the 24B and 32B models.

The thinking configurations have the three plain greedy runs only.

Each blinded round was sampled by script with a recorded seed, and the sha256 of its case list was written down before review. Round 1 has every greedy verdict where Marin strict and the proposed grader disagreed, 100 random agreements, and 136 fixtures. Round 2 was drawn from sample0 after the fixes that round 1 prompted, so its responses were unseen by every reviewer.

Reviewers saw an opaque case id, the instruction id and text, the full prompt, and the response. They never saw the model, a checker verdict, or the other reviewer's labels. Each reviewer batch was a separate agent run on Claude Opus 5.5 at max effort; parts of four round-1 batches ran on earlier Claude models. Reviewer B read the cases in a different order from reviewer A. The prompt is results/blind_audit/reviewer_prompt_v1.md in the repository.

Error rates

A grader's error rate is the share of the round's population verdicts that contradict the reviewers' consensus. It is estimated per stratum from the cases with a consensus, and weighted by the stratum's population:

from datasets import load_dataset

rows = load_dataset("xiaoyuanwang/ifbench-audit", "blind_review", split="train")
strata = load_dataset("xiaoyuanwang/ifbench-audit", "blind_review_strata", split="train")

def error_rate(round_, grader):
    sizes = {s["stratum"]: s["population"] for s in strata if s["round"] == round_ and s["population"]}
    rate = 0.0
    for stratum, size in sizes.items():
        decided = [r for r in rows if r["round"] == round_ and r["stratum"] == stratum and r["consensus"]]
        wrong = sum((r["consensus"] == "pass") != r[grader] for r in decided)
        rate += size * wrong / len(decided)
    return rate / sum(sizes.values())

print(error_rate(2, "marin_strict"), error_rate(2, "fixes_only"))  # 0.0607 0.0318

	Round 1	Round 2, held out
Cohen's kappa between reviewers	0.94	0.96
Marin strict	7.6% [4.7, 11.3]	6.1% [3.9, 10.1]
Fixes only	4.9% [2.1, 8.6]	3.2% [1.0, 7.2]
Proposed	4.2% [1.3, 7.9]	3.3% [1.1, 7.3]

The intervals come from 2,000 bootstrap resamples within strata.

Limitations

• Both reviewers are the same language model. They agree closely, and they may share blind spots.
• The sandbox did not enforce the reviewers' independence; the reviewers were instructed to keep it.
• The audit's author wrote the fixtures and their labels. In round 1, the reviewers' consensus contradicted 6 of the 131 fixture labels it decided.
• The decisions are open. A label that depends on one follows the reading stated in the row.

Licenses

• Fixtures, findings, decisions, and the reviewers' labels: Apache 2.0, as in the repository.
• Prompts from allenai/IFBench_test (in blind_review, and in some fixtures' arguments): ODC-BY-1.0, for research and educational use under Ai2's Responsible Use Guidelines.
• Model responses (in blind_review): each model's license. Apache 2.0 for Qwen2.5-7B/14B/32B-Instruct, Qwen3-8B/14B/32B, OLMo-2 (7B, 13B, 32B), marin-8b-instruct, Mistral-7B-Instruct-v0.3, and Mistral-Small-24B-Instruct-2501; MIT for phi-4; the Llama 3.1 Community License for Llama-3.1-Tulu-3-8B.