Add README describing top5 / last3 groupings
Browse files
README.md
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
tags:
|
| 4 |
+
- pi05
|
| 5 |
+
- lora
|
| 6 |
+
- robotics
|
| 7 |
+
- fruit-swap
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
# pi05 LoRA finetune — Fruit Swap
|
| 11 |
+
|
| 12 |
+
LoRA-finetuned `lerobot/pi05_base` on the FruitSwapEasy task.
|
| 13 |
+
Run: `19.55.32_train_pi05_fruit_swap_lora_70_20_10_fruit_swap_pi05` (split 70/20/10).
|
| 14 |
+
|
| 15 |
+
## Checkpoint groups
|
| 16 |
+
|
| 17 |
+
**Top 5 (mixed sort: phase1 desc, val asc):**
|
| 18 |
+
|
| 19 |
+
| Rank | Epoch | phase1 | val | filename |
|
| 20 |
+
|------|-------|--------|--------|---------------------------------------------------------|
|
| 21 |
+
| 1 | 44 | 0.30 | 0.0434 | epoch=0044-succ=0.00-phase1=0.30-val=0.0434.ckpt |
|
| 22 |
+
| 2 | 28 | 0.20 | 0.0386 | epoch=0028-succ=0.00-phase1=0.20-val=0.0386.ckpt |
|
| 23 |
+
| 3 | 40 | 0.20 | 0.0388 | epoch=0040-succ=0.00-phase1=0.20-val=0.0388.ckpt |
|
| 24 |
+
| 4 | 38 | 0.20 | 0.0390 | epoch=0038-succ=0.00-phase1=0.20-val=0.0390.ckpt |
|
| 25 |
+
| 5 | 48 | 0.20 | 0.0423 | epoch=0048-succ=0.00-phase1=0.20-val=0.0423.ckpt |
|
| 26 |
+
|
| 27 |
+
**Last 3 (by epoch among available ckpts):**
|
| 28 |
+
|
| 29 |
+
| Epoch | phase1 | val |
|
| 30 |
+
|-------|--------|--------|
|
| 31 |
+
| 48 | 0.20 | 0.0423 |
|
| 32 |
+
| 46 | 0.10 | 0.0441 |
|
| 33 |
+
| 44 | 0.30 | 0.0434 |
|
| 34 |
+
|
| 35 |
+
`succ` (eval rollout success) was 0.00 for all available ckpts at upload time;
|
| 36 |
+
`phase1` is the phase-1 sub-score, `val` is validation loss.
|
| 37 |
+
|
| 38 |
+
Note: epoch 22 was present locally but not in top5 or last3, so not included.
|