README.md

---
pretty_name: LuminBench Nano ESMC Full Open Reservoir v2
license: cc-by-sa-4.0
size_categories:
  - 100M<n<1B
tags:
  - biology
  - protein
  - protein-language-model
  - esmc
  - decontaminated
  - parquet
---

# LuminBench Nano ESMC Full Open Reservoir v2

This is the complete decontaminated 70%-identity representative reservoir for
[`Lumin-Science/LuminBench-Nano-ESMC`](https://github.com/Lumin-Science/LuminBench-Nano-ESMC).
It is organized as immutable, SHA-ordered Parquet shards so each run can download
only the smallest deterministic prefix required by its training budget.

## License and source terms

Lumin Science's original database selection, arrangement, decontamination
ledger, packing, and metadata are offered under CC BY-SA 4.0. Third-party
sequence records retain their source terms:

| Path | Direct source | Governing terms |
|---|---|---|
| `train/uniref90/**`, `validation/uniref90/**` | UniRef90 2023_02 | [CC BY 4.0](https://www.uniprot.org/help/license) |
| `train/mgnify/**`, `validation/mgnify/**` | MGnify Protein DB 2023_02 | [EMBL-EBI Terms of Use](https://www.ebi.ac.uk/about/terms-of-use/) plus applicable original-owner rights; not relicensed by Lumin Science |
| `train/omg_img/**`, `validation/omg_img/**` | JGI/IMG records from `tattabio/OMG` | [CC BY-SA 4.0](https://huggingface.co/datasets/tattabio/OMG) |

Redistribution notes and modifications are preserved in
`LICENSE_AND_ATTRIBUTION.md` and `SOURCE_PROVENANCE.json`.

## Decontamination

The exclusion union was built from the exact source data used by these
evaluations:

| Evaluation data/source | Corresponding evaluation task |
|---|---|
| [RCSB Protein Data Bank](https://www.rcsb.org/) snapshot dated 2024-02-28, using the [ESMC paper](https://biohub.ai/papers/esm_protein.pdf) contact protocol | Long-range contact prediction, reported as precision at L (P@L) |
| [TAPE remote-homology data](https://s3.amazonaws.com/songlabdata/proteindata/data_pytorch/remote_homology.tar.gz), derived from SCOP fold classes | Remote-homology fold classification |
| [TAPE secondary-structure data](https://s3.amazonaws.com/songlabdata/proteindata/data_pytorch/secondary_structure.tar.gz), with CB513 as the primary test set | Residue-level three-class secondary-structure prediction |
| [TorchDrug EnzymeCommission](https://zenodo.org/records/6622158/files/EnzymeCommission.zip) sequence adaptation | Multilabel Enzyme Commission function prediction; retained as a quarantined diagnostic |
| [DeepLoc 2.0 official five partitions](https://github.com/teevee112/DeepLoc-2.0/blob/8c4e712822e63aa67990d1ef4d78ad842e644120/data_files/multisub_5_partitions_unique.csv) | Multilabel subcellular-localization prediction |
| [PEER HumanPPI](https://miladeepgraphlearningproteindata.s3.us-east-2.amazonaws.com/ppidata/human_ppi.zip) | Human protein-protein-interaction prediction; retained as a quarantined diagnostic |
| [FLIP2 Hydrophobic Core low-to-high split](https://flip.protein.properties/assets/splits/hydro/low_to_high.csv.gz) | Protein-fitness regression under an engineering distribution shift |
| [CATH v4.4 S20 domains and classifications](https://download.cathdb.info/cath/releases/all-releases/v4_4_0/) | CATH remote structural-domain retrieval |
| [CAFA5 input bundle](https://zenodo.org/records/10951709) and [official final evaluation](https://zenodo.org/records/20186533) | Molecular-function transfer on the MF no-knowledge, 30%-identity hard set; blocked for scoring but still protected |
| [PRING](https://github.com/SophieSarceau/PRING/tree/edbdc0190e99c492337a9b3ddca6ebc2c6441924) | Human PPI prediction with node- and 30%-cluster-disjoint splits |
| [FLIP2 public split collection](https://flip.protein.properties/) | Multi-landscape protein-engineering shift evaluation |
| [RosettaCommons MegaScale](https://huggingface.co/datasets/RosettaCommons/MegaScale) | Family-held-out mutation-stability / delta-delta-G prediction |
| [CAID2 and CAID3 Disorder-PDB references](https://caid.idpcentral.org/challenge/results) | Temporal residue-level intrinsic-disorder prediction |

All sequences used to construct the listed task populations are included in the
protected union, including task fitting, validation, retrieval gallery, final
test, and the blocked CAFA candidate population.
The full representative reservoir is screened with MMseqs2 at 30% sequence
identity and 80% query plus target coverage. Exact SHA-256 matches are excluded
independently. The released validation union is excluded from every training
source arm.

## Download only what a run needs

```bash
git clone https://github.com/Lumin-Science/LuminBench-Nano-ESMC
cd LuminBench-Nano-ESMC
uv sync --frozen
uv run --frozen python scripts/download_data.py \
  --repo-id LuminScience/LuminBench-Nano-ESMC \
  --revision <immutable-release-commit> \
  --training-samples 5376000 \
  --cache-root data/cache/full-open-v2 \
  --output-root data/processed/run-prefix
```

The command fetches `manifest.json`, converts the requested total sample count
to per-source requirements using the 36:11:54 mixture, downloads the minimum
whole-shard prefix for each source plus every validation shard, verifies the
checksums, and materializes the existing mmap training layout. Training is local;
it does not make row-level network requests.

## Parquet schema

| Column | Type | Meaning |
|---|---|---|
| `sequence` | string | normalized amino-acid sequence |
| `sha256` | string | SHA-256 of the ASCII sequence |
| `length` | int32 | residue count |

See the GitHub processing report and `manifest.json` for full lineage, measured
counts, rejection accounting, shard hashes, and the complete build recipe.

## Verified release measurements

- Training representatives: **665,970,495**
- Training residues: **151,304,238,405**
- Training Parquet shards: **565**
- Validation representatives: **12,288**
- Validation residues: **3,182,651**
- Validation Parquet shards: **3**
- Train plus validation compressed bytes: **109,661,312,410**
- Evaluation-query union: **317,000** sequences
- Homology-excluded representative digests: **73,826,953**
- Manifest SHA-256: `fe1ac0657085ab19fe6f56786006e9eb004ca66bc6c5b81dfd8e6bc3dcfda6ff`

<div class="ai">

## Fixed evaluation settings

</div>

<div class="ai">

The current MLM evaluation scores all **12,288 held-out validation proteins**, exactly 4,096 from each source, once at context 512. Loss is the mean of per-protein masked-token negative log-likelihoods, giving each source one third of the score. [MLM_VALIDATION_12288.json](evaluation/MLM_VALIDATION_12288.json) records the v2 protocol, immutable source-shard hashes and reproducibility checks.

</div>

<div class="ai">

Each protein's crop and 15% mask positions derive from mask seed 20260821 and its sequence SHA-256 using a separate CPU random generator. Batch size changes throughput without changing those crops or masks. The training and validation Parquet files are unchanged; the [earlier 4,096-protein contract](evaluation/MLM_VALIDATION_4096.json) remains available for reproducing historical scores. Re-evaluate old checkpoints under v2 before comparing their losses with the full-set results.

</div>

<div class="ai">

[Evaluation assets](evaluation/README.md) include the clean contact bundle with all 20,775 evaluation chains, the fixed probe splits and unchanged scoring code. Its metadata contains dataset attribution and no experiment-report links or machine paths.

</div>