---
license: mit
tags:
  - protein
  - structure
  - pretrain
  - tokens
---

# K-Fold Structure Pretrain Corpus

ELECTRA pretraining corpus: 83.6M ATLAS + PDB structures, tokenized with
our backbone (aminoaseed VQ-VAE, from
[StructTokenBench](https://github.com/KatarinaYuan/StructTokenBench)) +
full-atom (CHI VQ-VAE) tokenizers.

## Format

Single LMDB. Per-record value (pickle):

```python
{
    "seq": int64 [L],   # AA sequence tokens
    "bb":  int64 [L],   # backbone tokens
    "fa":  int64 [L],   # full-atom tokens
}
```

Keyed by structure ID.

## Why sharded

The native `data.mdb` is ~654 GB, which exceeds HF's 50 GB per-file hard
limit. We byte-split into 17 parts of ~40 GB each. Reassemble before opening:

```bash
hf download k-fold-structure/triprorep-pretrain --local-dir ./pretrain.lmdb
cd ./pretrain.lmdb
cat data.mdb.part_* > data.mdb && rm data.mdb.part_*
# Now ./pretrain.lmdb/ is a valid LMDB.
```

## Quickstart

```python
import lmdb, pickle
env = lmdb.open("./pretrain.lmdb", readonly=True, lock=False, readahead=False)
with env.begin() as txn:
    rec = pickle.loads(txn.get(b"<your-key>"))
    print(rec["seq"].shape, rec["bb"].shape, rec["fa"].shape)
```

## Train/valid/test split

This LMDB contains all 83.6M structures pooled together, with no per-split
sub-LMDBs. For ELECTRA training, point the dataloader's ``lmdb_dir`` at
the reassembled directory and set ``train_split`` / ``val_split`` to
your own filter logic (or symlink ``train.lmdb`` → this dir if you want
to skip a validation pass).

## License

MIT