---
library_name: transformers
tags:
  - biology
  - protein-language-model
  - marinfold
---

# MarinFold exp166 checkpoints

Final checkpoints from the MarinFold contacts-v1 1.5B amino-acid augmentation
experiment tracked in
[Open-Athena/MarinFold#166](https://github.com/Open-Athena/MarinFold/issues/166).
Loss is `eval/tokenized/contacts-v1-val/loss`.

Both checkpoint formats are retained:

- `hf/` contains a `Qwen3ForCausalLM` Hugging Face safetensors export with the
  contacts-v1 tokenizer colocated with the weights.
- `checkpoints/` contains the original Levanter OCDBT checkpoint for training
  restart or re-export.

## Checkpoint inventory

| Epochs | Initialization | Selection | W&B run | Loss | Step | Repository paths | Full GCS sources |
|---:|---|---|---|---:|---:|---|---|
| 8 | exp117 checkpoint | Final | `prot-exp166-cv1-aaaug-1_5b-e8-lr3p162e-3-wd0p1-bs128-exp117-init-us-east1` | 2.664179 | 35,679 | Levanter: [`checkpoints/step-35679`](./prot-exp166-cv1-aaaug-1_5b-e8-lr3p162e-3-wd0p1-bs128-exp117-init-us-east1/checkpoints/step-35679)<br>HF: [`hf/step-35679`](./prot-exp166-cv1-aaaug-1_5b-e8-lr3p162e-3-wd0p1-bs128-exp117-init-us-east1/hf/step-35679) | Levanter: `gs://marin-us-east1/prot-exp166-cv1-aaaug-1_5b-e8-lr3p162e-3-wd0p1-bs128-exp117-init-us-east1/2026.07.27.1/checkpoints/step-35679/`<br>HF: `gs://marin-us-east1/prot-exp166-cv1-aaaug-1_5b-e8-lr3p162e-3-wd0p1-bs128-exp117-init-us-east1/2026.07.27.1/hf/step-35679/` |

The training target was 35,680 steps; Levanter's zero-indexed final checkpoint
is `step-35679`. The dated GCS layout component is intentionally omitted from
repository paths.