MutaBench data

Data companion to MutaBench: frozen biomolecular benchmark inputs, independent evaluator labels and prepared evidence. The code repository contains the evaluation interface, protocol implementations, reference baselines, saved metrics and figure packets.

Contents

Layer	Contents
tables/queries.parquet	61,688 label-free frozen queries from 217 protein and 32 RNA assays
tables/labels.parquet	The matching independent labels; load only in the evaluator
tables/tasks.json	Task identities, explicit label direction, metric and frozen candidate order
tables/labels.csv	Independent labels for the MutaBench CSV evaluator
archives/data-*.tar.gz	Source/canonical data, zero-shot inputs, few-shot and multiround data layers, MSA/structure evidence and navigation records
FILES.jsonl	Every packaged path, byte size, SHA256, archive index and group
RELEASE.json	Final archive/table checksums and release statistics
PLAN.json	Exact packaging scope and exclusions

Archive groups distinguish complex_sources, singlechain_sources, admet, zero_shot_inputs, raw_evidence, fewshot_multiround and navigation. Few-shot data include support/query splits and independent query evaluators. Multiround data include public candidate pools, initial examples and evaluation files. Some original filenames contain private_evaluator to enforce inference isolation of public benchmark labels.

The 249 registered single-chain assays are a source-pool count, not the denominator of every completed-model comparison or paper figure. Source eligibility, available evidence, admitted tasks, completed predictions and each paired cohort remain distinct.

Multiple frozen historical input versions are retained for provenance. They are not interchangeable releases. Apply zeroshot/EVALUATION_EXCLUSIONS.json and the selected input version's manifest when reproducing a current comparison; superseded prompts and excluded assays must not be silently reused. Canonical N250/N10 pools describe eligibility pools, not an already sampled test query.

Loading tables

from datasets import load_dataset

queries = load_dataset("yanjiehuang/MutaBench", "queries", split="test",
                       revision="<PINNED_DATA_COMMIT>")
# Evaluator process only:
labels = load_dataset("yanjiehuang/MutaBench", "evaluator_labels", split="test",
                      revision="<PINNED_DATA_COMMIT>")

Use the immutable revision recorded in the code repository's benchmark/data_release.json. Authentication is required while the repository is private. Larger predicted scores must follow each declared task's target direction. Independent query labels must not enter model prompts, retrieval, feature preparation or method-selection inputs.

Archive reproduction

Use the code repository's scripts/download_data.py to download the pinned release and verify archive/table hashes, then scripts/restore_data.py to safely extract into a fresh directory and verify every restored file. The archives use pipeline-relative paths. Frozen historical files retain their original hashes, including legacy path strings; they are not silently rewritten to claim that every old runner is portable.

Private cohorts, credentials, model weights, installed environments, provider responses and retry logs are excluded. Large optional upstream archives and historically missing references are listed separately. This release distributes the selected public benchmark data layers, not every research-server file or every historical experiment's raw API transcript.

Source terms and limitations

See SOURCES.md. Data retain source-specific terms, including the non-commercial restrictions on AbAgym-derived content. The code repository's MIT license does not relicense these data. Original task context, assay identity, coordinate mappings, missing features, score directions and provenance remain part of the evaluation contract. These data are for nonclinical computational evaluation of existing measurements.

No new model inference or experimental sequence construction was performed to create this distribution. Saved best-arm comparisons in the code repository are descriptive selection on their displayed cohorts, not held-out method selection.