Calpha / training_data /README.md
liuyong9975's picture
Calpha training data: README.md
b661793 verified
|
Raw History Blame Contribute Delete
2.23 kB

Calpha training data

Two data sets, packed as split tar archives. Join the parts before unpacking:

cat structure/targets_pdb.tar.part-* | tar -xf -
cat structure/tokens_pdb.tar.part-*  | tar -xf -
cat language/v5_vlm.tar.part-*       | tar -xf -
sha256sum -c SHA256SUMS

structure/ — 84,658 PDB chains (stage 1 and stage 2)

Protein chains from the Protein Data Bank in the input format of SimpleFold (derived from Boltz).

  • targets_pdb (12 GB): per-chain structure records (structures/*.npz, records/*.json) and manifest.json. Built from 98,254 PDB chains.
  • tokens_pdb (32 GB): the same chains tokenised and featurised for training, filtered to the 84,658 protein-only chains actually used (manifest.json; manifest.all.json lists all).

Both stages crop chains during training (256 residues; 512 in the last phase of stage 2). Built with training/data/simplefold_targets.py of the code repository.

language/ — stage-1 language data (v5_vlm, 33 GB)

Pre-tokenised chat-format samples, five epochs of about 547M tokens each (2.74B tokens in total; stage 1 consumed 2B). Per epoch, by tokens:

Source Share Content
protein tasks 75% inverse folding, secondary structure, infilling, contact prediction, written with amino-acid and structure-related tokens; sequences and labels derived from PDB and AlphaFold DB structures
fineweb_edu 17% general English text from the FineWeb-Edu corpus
mot_reasoning 8% general reasoning traces

Files: train/epoch<k>-<shard>.{jsonl,tok.npy,idx.npy,meta.parquet} (text, token ids, offsets, per-sample metadata), val/, tokenizer/ (Qwen3.5 tokenizer with the added tokens), summary.json. Homologs of the FoldBench-334 targets were removed from the protein samples.

Sources and terms

The structure data and the protein tasks derive from the Protein Data Bank (CC0) and the AlphaFold Protein Structure Database (CC BY 4.0). The general-text portion is redistributed from third-party corpora in tokenised and text form and remains under the terms of those corpora; if you reuse it, follow their licenses. The data is provided for research.