Moomboh/ESMC-300M-mutafitup

Multi-task LoRA fine-tuned ONNX models derived from ESM-C 300M by EvolutionaryScale.

Built with ESM.

ONNX Models

Model	Section	Tasks	Variant
ESMC-300M-mutafitup-accgrad-all-r4-best-overall	accgrad_lora	disorder, gpsite_atp, gpsite_ca, ... (16 total)	best_overall
ESMC-300M-mutafitup-align-all-r4-best-overall	align_lora	disorder, gpsite_atp, gpsite_ca, ... (16 total)	best_overall

Each ONNX model directory contains:

• model.onnx -- merged ONNX model (LoRA weights folded into backbone)
• export_metadata.json -- task configuration and preprocessing settings
• normalization_stats.json -- per-task normalization statistics
• tokenizer/ -- HuggingFace tokenizer files
• history.json -- training history (per-epoch metrics)
• best_checkpoints.json -- checkpoint selection metadata

PyTorch Checkpoints

The checkpoints/ directory contains minimal trainable-parameter PyTorch checkpoints for all training runs (45 runs across 4 training sections). These checkpoints contain only the parameters that were updated during fine-tuning (LoRA adapters and task heads), not the frozen backbone weights.

Each run directory (checkpoints/{section}/{run}/) contains:

• history.json -- training history
• best_checkpoints.json -- checkpoint selection metadata
• best_overall_model/model.pt -- best checkpoint by overall metric
• best_loss_overall_model/model.pt -- best checkpoint by overall loss
• best_task_models/{task}/model.pt -- best checkpoint per task metric
• best_loss_task_models/{task}/model.pt -- best checkpoint per task loss

To load a checkpoint, use MultitaskModel.load_trainable_weights() from the mutafitup training library.

License

The ESMC 300M base model is licensed under the EvolutionaryScale Cambrian Open License Agreement.

Fine-tuning code and pipeline are licensed under the MIT License.

See NOTICE for full attribution details.