---
license: mit
tags:
- robotics
- libero
- oat
- blockwise-decoding
---

# Blockwise LIBERO Finetune Artifacts (Phase 2)

This repository stores artifacts from the **LIBERO Phase-2 tail-decoder fit** for Blockwise-OAT.

## Fit Summary

- Resume checkpoint: `checkpoints/original_oat_tail_p4_r1.pt`
- Requested epochs: `30`
- Completed epochs: `10` (early stop triggered)
- Best validation loss: `2.7337` at epoch `1`
- Final epoch loss/val: `2.7343 / 2.7466`
- Tail/AR parameter ratio: `0.901` (`4,525,568 / 5,023,488`)
- Train/val split: `117,883 / 6,204` (`val_ratio=0.05`)

## Included Artifacts

- `checkpoints/original_oat_tail_p4_r1.pt` - resumed tail checkpoint after Phase-2 fit
- `checkpoints/original_oat_tail_p4_r1_best.pt` - best validation checkpoint
- `logs/libero_phase2_pipeline_B.log` - raw training/eval pipeline log from cluster
- `metrics/fit_summary_phase2.json` - parsed fit metrics snapshot
- `visualizations/tail_training_phase2_dashboard.png` - train/val loss dashboard

## Visualization

![Phase-2 tail training](visualizations/tail_training_phase2_dashboard.png)

## Notes

- This upload focuses on **fit artifacts** for LIBERO.
- Eval steps in the same pipeline log include later environment/runtime failures; those are kept in raw log for traceability.