---
pretty_name: AI4Sci FoldBench frozen data
tags:
- protein-folding
- foldbench
- ai4science
- int8
size_categories:
- 100K<n<1M
---

# AI4Sci FoldBench data

This repository hosts the **actual large data payloads**, not a copy of the
Harbor task. Task code, instructions, contract, reference implementations,
evaluation code and the download/extraction script live in
[T0-RSI/ai4sci-tasks](https://github.com/T0-RSI/ai4sci-tasks/tree/main/foldbench).

Release: `foldbench-100k-esmc-int8-v1`. The complete release is identified by
the immutable Hub commit pinned in the GitHub task's `download-source.json`;
do not treat an intermediate upload commit as a complete release.

## Contents

| Restored directory | Files | Bytes | Contents |
| --- | ---: | ---: | --- |
| `manifests/` | 3 | 291,989,362 | Frozen 100K logical training records and supporting manifests |
| `objects/` | 90,000 | 7,994,107,456 | Deduplicated training structure NPZ files |
| `features/` | 91,366 | 61,910,892,072 | Raw final-layer ESMC-6B features, groupwise INT8 |
| `foldbench/` | 4,623 | 2,116,164,115 | Complete native evaluation suite, ground truth and 1,568 feature files |
| `models/` | 3 | 755,429,835 | ESMFold2-Fast reference weights, config and upstream model card |
| **Total** | **185,995** | **73,068,582,840** | **73.1 GB decimal / 68.0 GiB** |

The evaluation suite contains 1,522 requests and 1,823 target rows across
nine tracks. This is the existing frozen task data, repackaged without
changing file contents. The feature identity is
`esmc6b-final-ln-int8-g64-45b0fa5d-ptsdpa-v1`: INT8 values `[L,2560]` and
BF16 scales `[L,40]`, not an ESMFold2-specific learned projection.

There are **no ESMC-6B model weights**, FP16/BF16 feature masters, MSA payloads,
experiment logs, credentials, or Harbor source files in this dataset.
ESMFold2-Fast weights are **reference-only**: the task forbids submitting
them or using them to initialize the trainable folding model. Features are
fixed inputs, not trainable submitted model parameters. See the GitHub
task contract for the complete research and evaluation rules.

## Download and restore

The payload is stored as 75 uncompressed, byte-preserving tar shards, with
approximately 1 GB of data per shard. `release-manifest.json` records each
archive's SHA-256, size and file count. `payload-files.jsonl` records the
SHA-256 and size of every restored file. The GitHub task pins both the Hub
commit and the archive-manifest checksum.

Use `foldbench/environment/data/materialize.sh DESTINATION` from the
GitHub task (with Python 3.11+ and the `hf` CLI installed). It downloads
individual pinned shards, verifies archive and payload hashes, restores
the original directory layout, and checks the scientific release metadata.
Successfully extracted downloaded shards are removed by default; source
data are never removed. `--keep-archives` retains the additional archive
copy. `--verify-only` rechecks an already materialized release offline.

## Provenance and rights

This is an AI4Sci task data bundle, not an official upstream release.
The original FoldBench benchmark, structure data and reference weights
retain their respective upstream attribution and terms. In particular,
the ESMFold2-Fast upstream model card and its third-party-notice link are
included unchanged alongside the reference weights. No blanket relicensing
of these third-party assets is claimed here. Feature files retain their
source model revision and sequence provenance in their safetensors metadata.