admin/lerobot
1
0
Fork 0
mirror of https://github.com/huggingface/lerobot.git synced 2026-07-29 20:49:42 +00:00
Code Issues Packages Projects Releases Wiki Activity

feat(policiy): fine tune vision encoder refactor completness

Browse Source
This commit is contained in:
Steven Palma
2026-07-29 18:35:24 +02:00
parent 9e30807eeb
commit 0371e99117
5 changed files with 66 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
AGENT_GUIDE.md
+3 -2
View File
@@ -321,10 +321,11 @@ SmolVLA ships with `freeze_vision_encoder=True`. Unfreezing usually **improves p
```bash
lerobot-train ... --policy.type=smolvla \
--policy.freeze_vision_encoder=false \
--policy.train_expert_only=false
--policy.fine_tune_vision_encoder=true
```
This selectively trains the vision encoder and connector while leaving the language model frozen. Their learning rate defaults to `0.1 × optimizer_lr`; adjust it with `--policy.vision_encoder_lr_multiplier` if needed.
### 7.7 Signals to stop / keep going
- Train loss plateaus → stop, save a Hub checkpoint.