BBH
Harbor adapterBIG-Bench Hard evaluates advanced reasoning capabilities across 23 challenging tasks using over 6,500 English samples.
Verified100%Knowledge & Reasoning
- Samples
- 6,511
- Released
- 2022
- Scores
- 59
FoldBench
FoldingCorpus / Fold2Reason Benchmark
A protein-derived question-answer dataset and post-training recipe evaluating general reasoning transfer from structural supervision across spatial, graph, scientific, and general benchmarks.
How to run
ltb · foldbench
$ npm i -g lettokenburn-cli$ ltb bench info foldbench
External harness. No Harbor adapter has been generated for this benchmark yet. `ltb bench info` prints the known links.
Runs happen on your machine with your own subscription; only the score summary is pushed. CLI docs
Scores
2 sourced entries| Harness | Source | Confidence | Link | |||
|---|---|---|---|---|---|---|
| InternVL3.5-8B-Instruct | 0.2 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
| qwen3.5-9B | 0.2 | N/A | Independent | medium | Sep 30, 2026 | arxiv.org |
Every number links to where it was published. Low-confidence rows are muted and need per-cell verification.
Reproduced by the community
All runsLoading community runs…
About
Content & Construction: The benchmark utilizes FoldingCorpus, a dataset constructed by applying 12 deterministic structural operators to the OpenFold monomer short-protein collection. It generates question-answer pairs verified against source coordinates in the Protein Data Bank (wwPDB). FoldBench evaluates models post-trained with this data via discrete language head supervision and continuous geometry decoding.
Scoring & Findings: Evaluation measures performance on 10 general-purpose benchmarks spanning spatial, graph, scientific, and reasoning categories. The primary metric is macro-average accuracy. Key findings show that training on protein structures improves broad reasoning capabilities (raising macro-accuracy from ~45% to ~48%), with positive gains across text, image, and video tasks compared to controls.