yanjiehuang/MutaBench-results

MutaBench — Experiment results

Eight main data categories · Supporting resources · Experiment results · Table guide

Public page; manually approved access is required for all data. Existing approved main-repository users were granted matching access at publication. Future approvals are separate. This repository presents the same scientific data and recorded runs with clearer navigation. Source columns and identifiers are unchanged; original assets remain accessible through the approved main repository.

Source revision: 5ad6c7319b454327725db09ad4bb16d1cca0308d. The full original files remain in MutaBench.

Choose a table

Table	Plain-language name	What it contains
few_shot_scores	Final experiment metrics	Metrics for 1,896 final ordinary and thinking experiment conditions, linked to queries, predictions and evaluator files.
few_shot_predictions	Final per-sample predictions	474,000 predictions in original dataset-ID namespace; join by protocol/case/model/effort and variant ID.
few_shot_queries	Few Shot Queries	26,500 unique query-file rows for the final ordinary/thinking cohorts; no query labels.
few_shot_reference_labels	Few Shot Reference Labels	Independent final-query labels; keep out of predictor inputs.
zero_shot_scores	Zero Shot Scores	315 newly recomputed results from three completed matched-zero model chains; separate from older paper metrics.
zero_shot_predictions	Zero Shot Predictions	78,750 matched-zero predictions on the frozen 105-assay cohort.
zero_shot_sampling	Zero-shot sampling blocks	Fixed sample blocks, hashes and model protocol for the matched zero-shot runs. Each row records a fixed sample block, hashes and model protocol.
higher_order_conditions	Higher Order Conditions	639 stored strict higher-order conditions; completion and metric availability are explicitly distinct.
higher_order_predictions	Higher Order Predictions	76,500 stored HIGH strict-order predictions; retain episode and order.
higher_order_sampling	Higher Order Sampling	3,060 selected HIGH strict-order blocks with source provenance.
paper_cohorts	Paper assay cohorts	Assays included in the paper cohort, with common and model-specific membership recorded separately.
historical_scores	Historical Scores	438,144 versioned source metric rows; select a source/protocol before comparison.
score_provenance	Score Provenance	285 original paper/multiround metric tables and their exact hashes.
support_query_examples	Support Query Examples	Original frozen supports, queries and pools; role is retained or inferred from source packaging.
experiment_collections	Experiment Collections	Published input collections and their available information modalities. Completion state is recorded separately for each experiment.
run_inventory	Run Inventory	Reviewed run inventory with recorded states and source counts; use its recorded state to identify completed runs.

Rows are file records, alignment sequences, array blocks, examples, or experimental conditions depending on the table. Each table uses its own row unit. Select an experiment/version before comparing scores. Keep query ground truth for independent evaluation.