Instructions to use open-athena/marinfold-exp146 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use open-athena/marinfold-exp146 with Transformers:
# Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("open-athena/marinfold-exp146", device_map="auto") - Notebooks
- Google Colab
- Kaggle
MarinFold exp146 checkpoints
Selected checkpoints from the MarinFold contacts-v1 3B experiment tracked in
Open-Athena/MarinFold#146.
Loss is eval/tokenized/contacts-v1-val/loss, and only finished W&B runs were
considered.
Both checkpoint formats are retained:
hf/contains aQwen3ForCausalLMHugging 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.
“Early-stop” denotes the lowest loss at an evaluation step for which a permanent Levanter checkpoint exists.
Checkpoint inventory
| Epochs | Selection | W&B run | Loss | Step | Repository paths | Full GCS sources |
|---|---|---|---|---|---|---|
| 2 | Final + checkpoint-aligned early-stop winner | prot-exp146-cv1-s01-3b-e2-lr3p162e-3-wd0p2-bs64-us-east1 |
2.926841 | 17,839 | Levanter: checkpoints/step-17839HF: hf/step-17839 |
Levanter: gs://marin-us-east1/checkpoints/protein/prot-exp146-cv1-s01-3b-e2-lr3p162e-3-wd0p2-bs64-us-east1/2026.07.23.01/checkpoints/step-17839/HF: gs://marin-us-east1/checkpoints/protein/prot-exp146-cv1-s01-3b-e2-lr3p162e-3-wd0p2-bs64-us-east1/2026.07.23.01/hf/step-17839/ |
| 8 | Final + checkpoint-aligned early-stop winner | prot-exp146-cv1-s01-3b-e8-lr3p162e-3-wd0p4-bs256-us-east1 |
2.702478 | 17,839 | Levanter: checkpoints/step-17839HF: hf/step-17839 |
Levanter: gs://marin-us-east1/checkpoints/protein/prot-exp146-cv1-s01-3b-e8-lr3p162e-3-wd0p4-bs256-us-east1/2026.07.23.01/checkpoints/step-17839/HF: gs://marin-us-east1/checkpoints/protein/prot-exp146-cv1-s01-3b-e8-lr3p162e-3-wd0p4-bs256-us-east1/2026.07.23.01/hf/step-17839/ |
- At 8 epochs, another run reaches a slightly lower raw evaluation loss of
2.701957at step 33,450, but no permanent checkpoint exists at that step. Its best checkpoint-aligned loss is2.708061at step 31,220, which is worse than the selected checkpoint (2.702478at step 17,839), so it is not included as an additional upload. - The dated GCS layout component is intentionally omitted from repository paths.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support