File size: 2,407 Bytes
c823ad4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
---
license: other
license_name: cambrian-open
license_link: https://huggingface.co/Moomboh/ESMC-300M-mutafitup/blob/main/CAMBRIAN_OPEN_LICENSE.md
tags:
  - protein-language-model
  - onnx
  - fine-tuning
  - multi-task
  - esm
---

# Moomboh/ESMC-300M-mutafitup

Multi-task LoRA fine-tuned ONNX models derived from
[ESM-C 300M](https://huggingface.co/EvolutionaryScale/esmc-300m-2024-12)
by [EvolutionaryScale](https://www.evolutionaryscale.ai/).

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](https://github.com/Moomboh/mutafitup) training library.

## License

The ESMC 300M base model is licensed under the
[EvolutionaryScale Cambrian Open License Agreement](CAMBRIAN_OPEN_LICENSE.md).

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

See [NOTICE](NOTICE) for full attribution details.