logica-v2 / README.md
super-dainiu's picture
Document actual v2 jobs and completion-only publication
17afad4 verified
|
Raw History Blame Contribute Delete
1.36 kB
---
library_name: pytorch
---
# LogiCA v2
This repository is for newly trained LogiCA v2 models only.
Original v1 weights remain in [Yale-CompBio/logica](https://huggingface.co/Yale-CompBio/logica).
| Protein backbone | Completed release destination |
| --- | --- |
| ESM-2 8M | `checkpoints/v2_20261110/8m/` |
| ESM-2 35M | `checkpoints/v2_20261110/35m/` |
| ESM-2 150M | `checkpoints/v2_20261110/150m/` |
| ESM-2 650M | `checkpoints/v2_20261110/650m/` |
These are planned destinations. A checkpoint directory becomes available only
when its full training run and release validation have completed.
Each completed release includes `best.pt`, the exact backbone/configuration,
model code, training completion record, data preparation metadata, and SHA-256 hashes.
The checkpoint is a v2 adapter overlay; instantiate the provided v2 architecture
with its pinned ESM-2 and SELFormer backbones before loading it.
All sizes use the same curated and decontaminated BindingDB v2 data, 40 epochs
with 100,000 sampled anchors per epoch, effective batch size 64, and 15 in-batch
negatives. All four sizes use four GPUs with B=16 per GPU and no gradient
accumulation. Runs save and continue across wall-time limits.
Small-model and 650M completion and publication are independent.
`release_plan.json` describes the submitted training plan, not completed results.