MarinFold exp199 checkpoints
Checkpoint artifacts from the MarinFold contacts-v1 AFDB/ESM optimization experiment tracked in Open-Athena/MarinFold#199.
Runs
| Run | Final eval/tokenized/contacts-v1-val/loss |
Levanter checkpoints | Hugging Face export |
|---|---|---|---|
prot-exp199-cv1-s01-m1-p03-aug-us-east1 |
3.0115306377 |
step-72599 | step-72599 |
prot-exp199-cv1-s01-m1-p03-base-us-east5 |
3.0074224472 |
step-72599 | step-72599 |
prot-exp199-cw-cv1-s02-m1-p06-aug |
2.9712009430 |
step-116160, step-145199 | step-145199 |
prot-exp199-cv1-cont-s03-m1-p03-srcbase-aug100-us-east1 |
2.9638261795 |
step-145199 | step-145199 |
prot-exp199-cw-cv1-p06-cool-s01 |
2.9396727085 |
— | step-290400 |
Formats
The checkpoints/ paths contain the original Levanter OCDBT full
state, including AdamW state for continued training. The hf/
paths contain Qwen3ForCausalLM safetensors exports and the
2,845-token contacts-v1 tokenizer.
The two original m1-p03 runs trained for 72,600 steps, making
step-72599 their zero-indexed final checkpoint. The CoreWeave
m1-p06 and continued srcbase-aug100 runs trained
through step-145199. Only final checkpoint formats are retained
for the continuation run.
p06-cool-s01 is a cooldown of the CoreWeave m1-p06
line. It resumed from step-261360 of
prot-exp199-cw-cv1-cont-s02-m1-p06-srcaug-aug100 and decayed the
learning rate linearly from 1e-3 to zero over 29,040 steps, ending
at step-290400. Only its Hugging Face export is retained.
We release our tools to the public with the intent that they are used safely, ethically, and responsibly. They are meant to contribute to the public's understanding of science, and aid in its advancement. By using this Marin model, you agree to comply with all applicable law.