# La-Proteina SafeTensors Weights

These are the pretrained weights for [La-Proteina](https://github.com/NVIDIA-Digital-Bio/la-proteina),
converted from the original PyTorch `.ckpt` checkpoints to SafeTensors format.

## Source

- **Original repository**: https://github.com/NVIDIA-Digital-Bio/la-proteina
- **Paper**: https://arxiv.org/abs/2507.09466
- **Project page**: https://research.nvidia.com/labs/genair/la-proteina/

## Files

### Score Networks

| File | Model | Description |
|------|-------|-------------|
| `LD1_ucond_notri_512.safetensors` | LD1 | Unconditional, no triangle update (up to 512 residues) |
| `LD2_ucond_tri_512.safetensors` | LD2 | Unconditional, with triangle update (up to 512 residues) |
| `LD3_ucond_notri_800.safetensors` | LD3 | Unconditional, long proteins (300-800 residues) |
| `LD4_motif_idx_aa.safetensors` | LD4 | Indexed motif scaffolding, all-atom (up to 256 residues) |
| `LD5_motif_idx_tip.safetensors` | LD5 | Indexed motif scaffolding, tip-atom (up to 256 residues) |
| `LD6_motif_uidx_aa.safetensors` | LD6 | Unindexed motif scaffolding, all-atom (up to 256 residues) |
| `LD7_motif_uidx_tip.safetensors` | LD7 | Unindexed motif scaffolding, tip-atom (up to 256 residues) |

### Decoders (Autoencoders)

| File | Model | Description |
|------|-------|-------------|
| `AE1_ucond_512.safetensors` | AE1 | Unconditional decoder (up to 512 residues) |
| `AE2_ucond_800.safetensors` | AE2 | Unconditional decoder (up to 800 residues) |
| `AE3_motif.safetensors` | AE3 | Motif scaffolding decoder |

## Conversion

These SafeTensors files were converted from the original PyTorch `.ckpt` checkpoints
using the `scripts/convert_weights_to_safetensors.py` script in the La-Proteina repository.
Weight tensors are stored as Float32. Only model weights are included (no optimizer state),
which is why the SafeTensors files are significantly smaller than the original checkpoints.

## License

These weights are licensed under the [NVIDIA Open Model License](LICENSE).
See the LICENSE file for full terms.