localgate/mmlu-pro-open-gemma-generations

MMLU-Pro Open — Gemma 4 E2B Generations & Evaluations

42,270 generated responses and 8,454 question-level evaluations, recording five attempts by google/gemma-4-E2B-it on converted, option-free MMLU-Pro questions. Labels come from a separate three-model judge panel. This release supports studying answer reliability and comparing multiple-choice and open-ended performance on the same questions.

Maintained by LocalGate for an AI & Sustainability research prototype. Last updated: 2026-09-07. The release inventory and source revisions are recorded in provenance/release.json.

Load the data

Install the Hugging Face datasets package and authenticate with an account that can access this private repository. token=True uses your saved credentials.

from datasets import load_dataset

repo = "localgate/mmlu-pro-open-gemma-generations"
evaluations = load_dataset(repo, "evaluations", split="test", token=True)
generations = load_dataset(repo, "generations", split="test", token=True)

evaluations is the default configuration. Parquet files need no custom loading script. Pin the Hub commit with revision="<commit SHA>" for a reproducible analysis.

Contents and joins

Split	evaluations: questions	generations: responses
train	5,851	29,255
validation	857	4,285
test	1,655	8,275
excluded_stem_leak	91	455
Total	8,454	42,270

Assignments follow the frozen LocalGate closed-corpus partition, restricted to the retained open questions. excluded_stem_leak preserves records excluded during duplicate-stem repair: do not use these rows for training or tuning. They remain available for provenance and descriptive analyses. The test split is for held-out evaluation; the all-corpus summary below is not a test-only result.

The study artifacts connect as follows:

1. Closed multiple-choice benchmark: original questions and five-attempt baseline labels.
2. Open-ended conversion: retained questions, reference answers, and conversion records.
3. This repository: generated responses and matched performance comparisons.
4. Judge panel: individual judgements and consensus labels.

Human conversion audits and judge calibration are documented separately in audit-results.

Join question-level records on integer question_id. A generation is uniquely identified by (question_id, attempt); its attempt equals the judge dataset's rep. The closed generation traces are a separate companion artifact; consult their card and revision when comparing run provenance.

Generation fields

Field	Type	Meaning
question_id	int32	Original MMLU-Pro identifier; join key across the collection.
attempt	int32	Repeat index, 0–4; corresponds to judge rep.
category	string	One of 14 source subject categories.
prompt	string	Converted question text supplied as the user message, without serialized chat-template tokens.
response_text	string	Complete stored completion, including any generated reasoning.
seed	int32	Engine seed: 5042 + attempt.
finish_reason	string	vLLM termination reason; all released records contain stop.
prompt_tokens	int32	Token count of the model input after chat-template application.
output_tokens	int32	Token count of the generated completion.

Evaluation fields

Field	Type	Meaning
question_id	int32	Question-level join key.
category	string	Source subject category.
list_style	bool	Whether the stripped original stem starts with “which of the following”, ignoring case; a narrow text heuristic.
open_correct_count	int32	Number of attempts graded correct by at least two of three judges; 0–5.
open_no_answer_count	int32	Attempts where any judge returned no_answer; 0–5. This can overlap with majority-correct attempts.
open_accuracy	float32	open_correct_count / 5, a proportion.
mc_correct_count	int32	Correct attempts on the same original multiple-choice question; 0–5.
mc_accuracy	float32	mc_correct_count / 5, a proportion.
accuracy_delta	float32	open_accuracy - mc_accuracy; multiply by 100 for percentage points.
performance_change	string	better, same, or worse, according to the two success counts.
prompt_text	string	Converted question; equals generations.prompt for that question.
reference_answer	string	Reference used by the judge panel.

There are always five attempts per question. The success count varies from 0 to 5; k=5 denotes the trial count, not the observed score. open_no_answer_count is neither a majority no-answer label nor a measured refusal rate.

Generation and grading

The run manifest records google/gemma-4-E2B-it, vLLM 0.25.1, PyTorch 2.11.0, and Transformers 5.14.1. Each repeat used a fresh engine with seed 5042–5046, temperature 1.0, top-p 0.95, top-k 64, and a maximum of 6,144 output tokens.

The exact message construction was:

messages = [{"role": "user", "content": open_question.strip()}]

vLLM's LLM.chat applied the model's default chat template. There was no system message, no few-shot example, and no custom stop sequence. The checked-in generator selects dtype="auto"; the preserved manifest does not establish a resolved precision or GPU model. See the generation protocol and original manifest for reproduction details and recorded limits.

Each response was graded by Ministral-3-14B, GPT-OSS-120B, and Qwen3-Next-80B through AWS Bedrock. Two or more match verdicts make an attempt correct; both no_match and no_answer map to zero. The panel card and protocol document the rubric; audit-results documents human calibration and its accepted threshold deviation. These are model-judged labels, not human ground truth for every response.

Descriptive results

Across all 8,454 retained questions, including excluded_stem_leak, mean five-attempt accuracy was 65.6% multiple-choice and 47.2% open-ended, a difference of −18.4 percentage points. The closed score here uses the matched subset; the full 12,032-question closed benchmark has a different aggregate. These summaries average question-level success proportions; they are not pass@5.

Subject	Questions	MC accuracy	Open accuracy	Difference (pp)
Math	1,159	87.0%	80.7%	−6.3
Computer science	259	75.7%	61.4%	−14.3
Economics	501	71.5%	56.6%	−14.9
Biology	384	79.4%	56.5%	−22.9
Business	640	77.7%	54.8%	−22.9
Physics	1,091	72.3%	50.3%	−21.9
Chemistry	961	71.4%	47.7%	−23.7
Psychology	408	63.6%	38.6%	−25.0
Health	407	63.2%	36.6%	−26.6
Philosophy	355	48.4%	34.9%	−13.5
Other	498	50.0%	31.6%	−18.4
Engineering	834	54.6%	26.4%	−28.2
Law	833	30.9%	26.0%	−5.0
History	124	47.3%	13.2%	−34.0
Overall	8,454	65.6%	47.2%	−18.4

Differences are computed before rounding, so displayed percentages may not subtract exactly.

Generation used 2,991,525 input tokens and 34,526,615 output tokens: 37,518,140 total, averaging 816.8 output tokens per response. All 42,270 records ended with stop; none recorded token-limit termination. This does not establish answer completeness or correctness. Verified run-level energy and emissions measurements are unavailable; token counts are not energy measurements.

Intended use and limitations

Use these artifacts for reproducible error analysis, comparison of answer formats, and research into model-specific difficulty. Do not treat a single model's five-attempt scores as universal question difficulty or as evidence of safe deployment.

The converted subset is selected rather than representative of all MMLU-Pro questions. Open and closed runs differ in prompt format, seeds, stopping behavior, and grading; the observed differences do not identify the causal effect of removing answer options. Reference errors, conversion errors, judge bias, and benchmark exposure may affect results. Generated text can contain factual errors or harmful content and is not professional advice.

Three groups of identical rewritten question stems still cross active partitions (six items), including one train/test pair. The frozen assignments are preserved, so the existing exclusion repair does not establish complete separation of rewritten stems.

Repeated attempts are clustered within questions. Analyses of uncertainty should respect that grouping rather than assume 42,270 independent questions. Keep excluded rows out of training and tuning, and do not tune against held-out test scores. Source/model revisions and hardware were not fully recorded; exact rerun equivalence cannot be promised.

License and citation

Repository metadata declares MIT, following the upstream benchmark. This declaration does not replace applicable upstream content or model terms; see MMLU-Pro and the Gemma model card. Report the dataset configuration and Hub commit alongside the citations.

@misc{localgate2026open_generations,
  author = {{LocalGate}},
  title = {MMLU-Pro Open: Gemma 4 E2B Generations and Evaluations},
  year = {2026},
  url = {https://huggingface.co/datasets/localgate/mmlu-pro-open-gemma-generations}
}
@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}
}