---
pretty_name: FoldingCorpus
license: mit
configs:
- config_name: default
  default: true
  data_files:
  - split: train
    path: corpus/train.jsonl
  - split: validation
    path: corpus/dev.jsonl
  - split: test
    path: corpus/frozen_test.jsonl
- config_name: protein_training
  data_files:
  - split: train
    path: protein_training/train.jsonl
  - split: validation
    path: protein_training/dev.jsonl
  - split: test
    path: protein_training/test.jsonl
---
# FoldingCorpus

Protein-derived supervised question-answer records for Fold2Reason. The default
configuration contains 12,000 train, 1,200 validation, and 1,200 frozen-test
question records from 1,000 / 100 / 100 proteins. The `protein_training`
configuration contains the associated BB4Q10 input/target records (1,000 / 100 / 100).
Audit and manifest JSON files are retained under `corpus/` but excluded from the
`load_dataset` split definitions.

```python
from datasets import load_dataset
corpus = load_dataset("GENTEL-Lab/FoldingCorpus")
proteins = load_dataset("GENTEL-Lab/FoldingCorpus", "protein_training")
```

Paper: https://huggingface.co/papers/2609.38879  
Code and reproduction instructions: https://github.com/GENTEL-lab/Fold2Reason

Source: project `dataset.zip`, SHA-256 `97ef38130fec0b7caaa323022d5b7bef620e7b26dfa42a1da4713066806dd90b`. The original release files
are unchanged. Exact per-file hashes and source paths are in `SOURCE_PROVENANCE.json`.
The protein records derive from a selected OpenFold high-confidence distillation
snapshot. The project owner confirmed redistribution for this public release;
the original acquisition record is documented as pending in the project release checklist.