localgate/mmlu-pro-gemma-traces

MMLU-Pro Gemma 4 E2B — Closed-Answer Generation Traces

The 60,160 raw completions behind localgate/mmlu-pro-closed: 12,032 MMLU-Pro multiple-choice questions, each answered five times by google/gemma-4-E2B-it. Each row preserves the rendered prompt, full generated output, extracted option letter, and harness score.

Release correction

This release replaces the earlier July run with the verified 12 August 2026 run. In the earlier release, attempts 1 and 4 had identical outputs on all 12,032 questions; the sampler seed had not reached vLLM. The corrected run passes distinct seeds 42–46 to vLLM, and no pair of repeats duplicates the entire corpus. Some individual outputs may still coincide by chance.

The replacement changes 60,131 generated outputs and 9,591 attempt scores. Do not combine earlier trace revisions with the corrected closed labels. Pin Hub revisions when reproducing a result. The current traces reproduce every closed correct_count and invalid_count exactly.

Load and join

Last updated: 2026-09-07. Authenticate with a Hugging Face account that can access this private dataset.

from datasets import load_dataset

traces = load_dataset("localgate/mmlu-pro-gemma-traces", split="train", token=True)
closed = load_dataset("localgate/mmlu-pro-closed", split="test", token=True)
test_ids = set(closed["question_id"])
test_traces = traces.filter(lambda row: row["question_id"] in test_ids)

The single Hub split, train, is an archive of all 60,160 attempts, including questions assigned to the study's validation and test partitions. It is not a training-only partition. Use the closed dataset's splits or provenance/split.json to select a study partition.

The unique row key is (question_id, attempt). question_id connects the study:

Dataset	Contents
Closed benchmark	Original questions, answer key, five-attempt scores, and study splits
Open benchmark	Option-free rewrites and conversion records
Open generations	Five completions per retained question and matched performance comparisons
Judge panel	Three-judge verdicts and open-answer consensus labels

These are closed-answer traces, with answer options in the prompt. The open generations are a separate experiment on the 8,454 retained questions.

Data fields

Field	Meaning
question_id	Integer MMLU-Pro identifier; join on this rather than question text.
attempt	Zero-based repeat, 0–4; maps to sampler seeds 42–46.
category	Subject recorded in the corrected evaluation run.
prompt	Exact rendered input, including chat template, instructions, question, and options.
reasoning	Complete generated text, including any working and final answer; not a validated derivation or a record of internal reasoning.
pick	Harness-extracted option letter, or the literal [invalid] when extraction failed.
answer	Correct option letter from MMLU-Pro.
correct	Harness exact-match score as a boolean; invalid extractions score false.
reasoning_chars	Python character count of the generated text.
doc_hash	Harness content hash of the source question document.
prompt_hash	Harness content hash of the rendered prompt.

Every question occurs once in each attempt. Rendered prompts, doc_hash, and prompt_hash are identical across a question's five attempts.

Generation and scoring

The run manifest is preserved in provenance/labels_manifest.json.

Setting	Recorded value
Model	google/gemma-4-E2B-it
Task / harness	mmlu_pro / lm-evaluation-harness 0.4.12
Backend	vLLM 0.25.1; PyTorch 2.11.0; Transformers 5.14.1
Hardware	One NVIDIA GeForce RTX 5090, 31.8 GB reported VRAM
Precision / context	dtype=auto; maximum model length 8,192
Prompting	Zero-shot; thinking mode disabled; task explicitly requests step-by-step text and a final option letter
Sampling	Temperature 1.0; top-p 0.95; top-k 64; seeds 42, 43, 44, 45, 46
Generation cap	6,144 tokens
Stop sequence	Question:

The complete input is available in each row's prompt. The harness requests a conclusion in the form the answer is (X) and scores its extracted letter against the authoritative answer key. The export preserves the harness score and extraction result rather than regrading the generated text.

Across the full benchmark, 36,478 / 60,160 attempts are correct (60.63%). There are 730 invalid extractions (1.21%); these may include correct answers written in an unrecognised format or outputs stopped before a final letter. Invalid extraction is not a measured refusal or truncation rate. The Question: stop can end a completion when that text appears in generated working. The raw sample records do not contain finish reasons, so a complete stop-versus-length breakdown cannot be reported.

The manifest records hardware and software versions, but not a resolved model revision; its Git SHA is unknown. No measured energy or carbon footprint is available for this run.

Uses and limitations

Use these traces to inspect extraction errors, compare attempts, and reproduce the closed success counts. They describe one model, task prompt, and sampling setup; they do not establish general question difficulty or correctness of the generated reasoning. Five attempts give a noisy estimate of per-question success probability.

The outputs are uncurated and can contain factual errors, repetition, or sensitive benchmark content. Inputs are public exam questions, not participant records. No comprehensive personal-information or harmful-content screening is claimed. Training-data contamination is unmeasured. Keep the study partitions intact when training or evaluating predictors from these records.

Source, licence, and citation

Questions and answer keys derive from TIGER-Lab/MMLU-Pro, which is MIT licensed. Consult the Gemma Terms of Use for the generating model's terms. Report the dataset's Hub revision alongside the upstream citation when using these corrected traces.

@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}
}