open-athena/marinfold-exp117

MarinFold exp117 checkpoints

Selected checkpoints from the MarinFold contacts-v1 1.5B experiment tracked in Open-Athena/MarinFold#117. Loss is eval/tokenized/contacts-v1-val/loss, and only finished W&B runs were considered.

Both checkpoint formats are retained when available:

• 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.

“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
8	Final + checkpoint-aligned early-stop winner	prot-exp117-cv1-s02-1_5b-e8-lr3p162e-3-wd0p2-bs64-europe-west4	2.713059	71,359	Levanter: checkpoints/step-71359 HF: hf/step-71359	Levanter: gs://marin-eu-west4/checkpoints/protein/prot-exp117-cv1-s02-1_5b-e8-lr3p162e-3-wd0p2-bs64-europe-west4/2026.07.13.02/checkpoints/step-71359/ HF: gs://marin-eu-west4/checkpoints/protein/prot-exp117-cv1-s02-1_5b-e8-lr3p162e-3-wd0p2-bs64-europe-west4/2026.07.13.02/hf/step-71359/
16	Final winner	prot-exp117-cv1-s02-1_5b-e16-lr3p162e-3-wd0p2-bs256-europe-west4	2.703709	35,679	Levanter: checkpoints/step-35679 HF: hf/step-35679	Levanter: gs://marin-eu-west4/checkpoints/protein/prot-exp117-cv1-s02-1_5b-e16-lr3p162e-3-wd0p2-bs256-europe-west4/2026.07.13.02/checkpoints/step-35679/ HF: gs://marin-eu-west4/checkpoints/protein/prot-exp117-cv1-s02-1_5b-e16-lr3p162e-3-wd0p2-bs256-europe-west4/2026.07.13.02/hf/step-35679/
16	Early-stop checkpoint winner	prot-exp117-cv1-s02-1_5b-e16-lr3p162e-3-wd0p2-bs256-europe-west4	2.696074	33,450	Levanter: checkpoints/step-33450 HF: not available at this step	Levanter: gs://marin-eu-west4/checkpoints/protein/prot-exp117-cv1-s02-1_5b-e16-lr3p162e-3-wd0p2-bs256-europe-west4/2026.07.13.02/checkpoints/step-33450/

• At 16 epochs, the same run wins both comparisons. Its earlier permanent checkpoint at step 33,450 has lower loss (2.696074) than its final checkpoint at step 35,679 (2.703709), so both checkpoints are retained.
• The dated GCS layout component is intentionally omitted from repository paths.