localgate/mmlu-pro-open

MMLU-Pro Open — Open-Ended Conversion

Last updated: 2026-09-07. This dataset contains 8,454 retained MMLU-Pro questions rewritten to be answerable without their multiple-choice options, their original forms, and the full record of 12,032 conversion decisions. Retention is a measured automated filter outcome; it is not a guarantee that every rewrite is unambiguous or correct.

The correct_count, picks, and attempt_correct fields in this repository describe Gemma's multiple-choice attempts. Evaluated free-text performance is available separately in Open Gemma generations.

This release updates the inherited multiple-choice labels to the August 12 rerun and aligns the retained questions to the study's frozen partition, including excluded_stem_leak. The conversion decisions and rewritten questions are unchanged. See release provenance and split assignments.

Load the data

The repository is private. Authenticate with a Hugging Face account that has access; token=True uses your configured credential.

from datasets import load_dataset

repo = "localgate/mmlu-pro-open"
questions = load_dataset(repo, "open_text", split="test", token=True)
controls = load_dataset(repo, "multiple_choice", split="test", token=True)
decisions = load_dataset(repo, "conversion_record", split="train", token=True)

row = questions[0]
prompt = row["question"]  # Do not include options or answer fields in an open prompt.
reference = row["answer_text"]

Use a Hub commit SHA as revision= when reporting results. conversion_record/train is a storage convention for the complete decision ledger, not a model-training partition.

How the study datasets connect

All four core datasets join on MMLU-Pro's question_id:

Dataset	Contents
MMLU-Pro Closed	Original questions and five-attempt Gemma multiple-choice scores
This dataset	Retained open questions, matched original questions, and conversion decisions
Open Gemma generations	Five free-text answers per retained question and evaluated performance
Judge panel	Three judges' bulk verdicts and consensus labels

The separate closed Gemma traces repository supplies the corrected August-run multiple-choice responses. Match release revisions when joining it to the inherited labels; open-protocol responses are in the separate open generations dataset.

Configurations and splits

• open_text (default): 8,454 retained questions. question contains the rewritten stem; answer choices remain in separate columns for reference and must not be sent to an open-ended answering model.
• multiple_choice: the same 8,454 questions, in the same splits, with their original stems. Other fields match open_text. This is the matched control for format comparisons.
• conversion_record: all 12,032 source questions and their conversion decisions, including the 3,578 rejected questions. Published as one train split.

The two retained-item configurations use the following assignments:

Split	Questions	Use
train	5,851	Fit models
validation	857	Select models and settings
test	1,655	Held-out evaluation
excluded_stem_leak	91	Provenance only; never train or tune on these rows
Total	8,454	All retained questions preserved

Assignments are inherited by question_id from the frozen closed partition. The excluded rows preserve questions removed from training or validation to address cross-partition stem overlap; they are not newly rejected conversions. Preserve this distinction when computing corpus counts. After source-stem exclusions, three groups of identical rewritten stems (six questions) still cross active partitions. The frozen study assignments are preserved and this residual overlap remains a limitation. The split repair also does not demonstrate absence of model-pretraining contamination or other semantic overlap.

Conversion method

The input is the 12,032-item test portion of TIGER-Lab/MMLU-Pro. Reference answers are the correct option texts. Surrounding whitespace is stripped from source text. deepseek:deepseek-v4-flash performed three separate passes at temperature 0:

1. Judge: read the question, options, and correct answer; decide whether the stem has that answer without the options.
2. Re-score: reconsider first-pass rejects on a 1–10 scale; readmit at score ≥ 5.
3. Rewrite: receive only the original stem, then write a standalone question or refuse. The rewrite pass receives neither answer choices nor the reference answer.

The prompt-and-schema fingerprint is 93057f756115. The protocol supplement contains the exact instructions and output schemas. Separating calls and withholding the answer during rewriting reduce opportunities for answer-driven rewrites; they do not prove that rewrites preserve meaning. Model verdicts are retained, including contested decisions.

Final outcome	Questions
Kept through first-pass acceptance	7,704
Kept after re-score	750
Dropped at re-score	3,562
Dropped at rewrite	16
Retained	8,454 (70.3%)

All 12,032 items reached a decision. Of the survivors, 2,566 retain the original stem unchanged. Per-pass recorded usage sums to 57,621,206 tokens; this is token accounting, not a measured energy or carbon footprint. Temperature 0 does not guarantee identical outputs on a later provider run; the published conversion ledger is the artifact of record.

Data fields

open_text and multiple_choice

Field	Meaning
question_id	Original integer MMLU-Pro identifier; join key across the study.
question	Rewritten stem in open_text; original stem in multiple_choice.
options	Original ordered options, retained for reference. Exclude them from open prompts.
category, src	Original subject and source benchmark identifier.
answer	Correct option letter.
answer_index	Zero-based correct-option index.
answer_text	Correct option text; the reference used for open-answer grading.
correct_count, k	Correct multiple-choice attempts and their denominator; k = 5.
picks	Five parsed multiple-choice selections in seed order 42–46; invalid extraction is retained as a marker.
attempt_correct	Five multiple-choice correctness booleans; sum equals correct_count.
invalid_count	Multiple-choice attempts with the harness extraction result [invalid]; included as incorrect in the count.
list_style	Whether the stripped original stem starts with which of the following, ignoring case.

conversion_record

Field	Meaning
question_id, question, category, list_style	Source identifier, original stem, subject, and text heuristic.
reference_answer	Original correct-option text.
convertible	Final retention decision. Filter on this field to recover the 8,454 survivors.
open_question	Retained rewrite; empty for rejected items.
stage	Stage associated with the final decision: judge, rescore, or rewrite.
score	Re-score value from 1 to 10; null when re-scoring did not run.
reason	Converter's explanation accompanying the final decision.
failure	Diagnostic failure tag; it does not determine membership.
judge_convertible, judge_reason, judge_failure	Original first-pass verdict, explanation, and diagnosis.
judge_tokens_in, judge_tokens_out	Recorded input/output tokens for the first pass.
rescore_tokens_in, rescore_tokens_out	Re-score tokens; null when that pass did not run.
rewrite_tokens_in, rewrite_tokens_out	Rewrite tokens; null when that pass did not run.
model, prompts	Converter identifier and prompt-and-schema fingerprint.

Failure tags describe none, open_set (many possible answers), underspecified_stem (missing information), negation (excludes alternatives), ranks_options (compares supplied choices), true_only_among_these (valid only relative to the choices), and option_relative (answer refers to options). Do not select failure == "none" instead of convertible. There are 142 retained items with a failure tag and 32 rejected items tagged none. First-pass and re-score opinions come from the same converter, not independent annotators.

Human audit

Three study authors completed conversion and rewrite audits, closed on September 1, 2026. Population-weighted disagreement was 21.6% on 248 conversion decisions; same-question error was 4.46% on 152 retained rewrites, narrowly passing the ≤5% gate. These are sample estimates, not certification of every row. The separate audit dataset contains the annotations, scoring details, sensitivity analysis, and known annotation limitations, alongside the later judge-calibration stage. Published conversion verdicts remain unchanged.

Uses and limitations

Use the retained questions for free-text evaluation and matched format comparisons, or use the full ledger to study conversion decisions. This benchmark does not measure intrinsic difficulty, general user-prompt performance, or unrestricted answer correctness.

• Selection changes the population. With August labels, mean multiple-choice accuracy is 65.626% for retained questions versus 60.635% over all 12,032 originals. Compare open and closed results on the same IDs; the matched config supports this. The difference alone does not explain why an item was retained or why a model answered it correctly.
• Reference answers can be incomplete. Option text may omit units or context supplied by the original alternatives. Rewrite fidelity does not independently validate every answer.
• Automated conversion remains fallible. The audit measures disagreement and rewrite drift; it does not certify each row. Re-score band 5 permits ambiguity under its own rubric.
• MC labels are noisy and model-specific. Five attempts are a coarse measurement under one prompting and extraction protocol. The closed card documents extraction failures.
• Content and coverage are inherited. Subject balance, possible source errors, and pretraining contamination remain limitations. Source questions, rewrites, and explanations may contain names or sensitive material; no comprehensive personal-information audit is claimed.

Licence and citation

The source MMLU-Pro card identifies its licence as MIT. This repository retains MIT metadata. Consult source licences and model terms for the respective materials. Cite MMLU-Pro and identify this dataset's Hub revision:

@article{wang2024mmlupro,
  title = {MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark},
  author = {Wang, Yubo and Ma, Xueguang and Zhang, Ge and others},
  journal = {arXiv preprint arXiv:2406.01574},
  year = {2024}
}

@misc{localgate2026open,
  author = {{LocalGate}},
  title = {MMLU-Pro Open: Open-Ended Conversion},
  year = {2026},
  url = {https://huggingface.co/datasets/localgate/mmlu-pro-open}
}

For corrections, open a dataset discussion with the affected question_id and revision, or use the LocalGate issue tracker.