[Release] Add DeXposure-FM v1.0 Model Checkpoints for DeFi Credit Exposure Forecasting
#1
by
losdwind
- opened
No description provided.
🚀 Initial Model Release: DeXposure-FM v1.0
Summary
This PR adds the complete DeXposure-FM model checkpoints for DeFi credit exposure network forecasting.
Files Added
| File | Size | Description |
|---|---|---|
dexposure-fm-h1.pt |
80MB | Fine-tuned model for 1-week horizon (AUPRC: 0.978) |
dexposure-fm-h4.pt |
79MB | Fine-tuned model for 4-week horizon (AUPRC: 0.973) |
dexposure-fm-h8-h12.pt |
79MB | Fine-tuned model for 8/12-week horizons (AUPRC: 0.967) |
graphpfn-frozen-all-horizons.pt |
79MB | Frozen GraphPFN baseline for all horizons |
README.md |
- | Model card with usage instructions |
config.json |
- | Model configuration |
metrics-*.json |
- | Evaluation metrics for each checkpoint |
Model Performance
| Horizon | AUPRC | AUROC | Edge MAE |
|---|---|---|---|
| h=1 week | 0.978 | 0.996 | 2.09 |
| h=4 weeks | 0.973 | 0.995 | 2.49 |
| h=8 weeks | 0.967 | 0.993 | 2.55 |
| h=12 weeks | 0.967 | 0.993 | 2.65 |
Checklist
- Model checkpoints uploaded
- README with usage instructions
- Model card metadata (YAML frontmatter)
- Evaluation metrics included
- License (Apache 2.0)
Please merge to make the models publicly available! 🎉
losdwind
changed pull request title from
Upload folder using huggingface_hub
to [Release] Add DeXposure-FM v1.0 Model Checkpoints for DeFi Credit Exposure Forecasting
📝 Update: README Revised to Match Paper Results
Changes in this update:
- ✅ Added comparison between paper results vs released checkpoints
- ✅ Added note explaining h1 model improvement (AUPRC 0.978 vs 0.972 in paper)
- ✅ Added ROLAND and Persistence baselines to evaluation tables
- ✅ Updated checkpoint table with "Note" column to clarify which match paper
Model Checkpoint Summary
| Checkpoint | Paper | Released | Status |
|---|---|---|---|
| h1 | AUPRC 0.972 | AUPRC 0.978 | ⬆️ Improved |
| h4 | AUPRC 0.973 | AUPRC 0.973 | ✅ Matches |
| h8 | AUPRC 0.967 | AUPRC 0.967 | ✅ Matches |
| h12 | AUPRC 0.967 | AUPRC 0.967 | ✅ Matches |
Note: The h1 checkpoint slightly outperforms paper results due to continued training improvements after submission. This demonstrates the robustness of the method.
Ready for merge! 🚀
losdwind
changed pull request status to
merged