Downloads · 30 days
13
21% of all-time downloads
zhoumiaosen/smolvla-libero10-full-finetune
smolvla-libero10-full-finetune is a robotics model from zhoumiaosen. Use it for the robotics task on the model card, and read the license before you ship it in a product. It is set up for lerobot. The card lists the license as apache-2.0.
A real fine-tune of lerobot/smolvlabase via LeRobot on the full LIBERO-Long (libero10) subset of HuggingFaceVLA/libero — 379 episodes across all 10 tasks, not a toy subset. Trained with the exact same dataset, episode…
Downloads · 30 days
13
21% of all-time downloads
All-time downloads
62
Public
Parameters
450M
907 MB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors907 MB · 100%
How the weights are stored.
BF16447M · 99%
From the Hugging Face model README
A real fine-tune of lerobot/smolvla_base
via LeRobot on the full LIBERO-Long (libero_10)
subset of HuggingFaceVLA/libero — 379 episodes across all 10 tasks, not a toy subset.
Trained with the exact same dataset, episode split, and step budget as our companion
GR00T-N1.5-3B fine-tune (zhoumiaosen/groot-n1p5-lora-libero10-full-finetune),
for a directly comparable fine-tuned-model result.
lerobot/smolvla_base (SmolVLM2-500M-Video-Instruct VLM backbone)HuggingFaceVLA/libero, episodes 0–378 (the 379 episodes belonging
to the 10 libero_10 tasks; the dataset's other 1,314 episodes, from
libero_spatial/libero_object/libero_goal/libero_90, were excluded).LeRobot's dataset chunking bundles many episodes per underlying storage
chunk for efficiency. Requesting --dataset.episodes=[0..378] correctly
identifies which chunks to load, but doesn't filter rows within those
chunks down to just the requested episodes — so of 244,370 loaded rows,
142,901 (58.5%) belonged to episodes we never asked for, causing dataloader
IndexErrors partway through training. Filtering dataset.hf_dataset down
to the requested episode set fixes this — but naively, dataset.hf_dataset.filter(lambda ex: ex["episode_index"] in requested)
silently drops every row: the dataset's set_transform converts
episode_index into a torch.Tensor even when only that one column is
requested, and PyTorch tensors hash by object identity rather than by
value, so tensor in {set of ints} always evaluates False. The fix casts
to a plain Python int first: int(ep_idx) in requested, combined with
input_columns=["episode_index"] to skip formatting the other (much
larger) columns during the filter, which also cut the filter step from
20+ minutes down to under a second.
Evaluated on the full libero_10 (LIBERO-Long) suite, 5 episodes per task
(50 total), same protocol as our GR00T fine-tune. Eval episodes use
different seeds than training, so this reflects generalization, not
memorization of the training set.
Overall: 4/50 = 8.0% success (vs. 0/50 = 0% for zero-shot SmolVLA, no fine-tuning, identical protocol)
Task (libero_10 id) | Description | Fine-tuned, 30k steps (5 ep) | Zero-shot (5 ep) |
|---|---|---|---|
| 0 | put both the alphabet soup and the tomato sauce in the basket | 0/5 | 0/5 |
| 1 | put both the cream cheese box and the butter in the basket | 0/5 | 0/5 |
| 2 | turn on the stove and put the moka pot on it | 1/5 | 0/5 |
| 3 | put the black bowl in the bottom drawer of the cabinet and close it | 1/5 | 0/5 |
| 4 | put the white mug on the left plate and put the yellow and white mug on the right plate | 0/5 | 0/5 |
| 5 | pick up the book and place it in the back compartment of the caddy | 1/5 | 0/5 |
| 6 | put the white mug on the plate and put the chocolate pudding to the right of the plate | 1/5 | 0/5 |
| 7 | put both the alphabet soup and the cream cheese box in the basket | 0/5 | 0/5 |
| 8 | put both moka pots on the stove | 0/5 | 0/5 |
| 9 | put the yellow and white mug in the microwave and close it | 0/5 | 0/5 |
On the 6-task subset (ids 2, 3, 4, 5, 6, 9) used elsewhere in our project, this checkpoint scores 4/30 = 13.3%, essentially matching our GR00T fine-tune's 4/30 = 13.3% on the same subset — two different VLA architectures, trained on the exact same 379 episodes for the exact same 30,000 steps, converging to a very similar generalization result.
Absolute success rates are still low, consistent with LIBERO-Long's long-horizon, multi-subgoal tasks needing substantially more than 379 training episodes to master. The meaningful result here is the direction: fine-tuning took this checkpoint from 0% to a real, non-trivial, non-memorized success rate on held-out-seed episodes — and did so without hitting any of the training instabilities our GR00T run needed to work around.
load_vlm_weights=true, no LoRA).