SmolVLA LIBERO Few-shot
Collection
SmolVLA LIBERO seen origin and Naive, LoRA, Replay, Frozen-Stats, and L2-SP few-shot checkpoints. • 4 items • Updated • 1
SmolVLA seen-expert on LIBERO-90 (100k)
Frozen seen-domain SmolVLA checkpoint after 100k imitation steps on
libero_90. This is the immutable origin for every target few-shot run in
the project. It has not been trained on the three held-out libero_goal
tasks.
Code: github.com/alexsuw/smolvla-libero-fewshot
Few-shot family (30 naive FT checkpoints):
alexsuw/smolvla-libero-fewshot-naive-baseline
Collection:
alexsuw/smolvla-libero-few-shot-6a8b009357482d2b4b9d3c2f
Research question
How many held-out expert demonstrations does SmolVLA need, after LIBERO-90 domain adaptation, to solve a new language-conditioned manipulation task?
Pinned sources
| Artifact | Link | Revision / id |
|---|---|---|
| Base policy | lerobot/smolvla_base | c83c3163b8ca9b7e67c509fffd9121e66cb96205 |
| SmolVLA | arXiv:2506.01844 | paper |
| Dataset | nvidia/LIBERO_LeRobot_v3 | e5907374380b8f96511957e6ba5582be52a1e179 |
| LIBERO | arXiv:2306.03310 · code | benchmark |
| LeRobot | huggingface/lerobot | d451fe4f1f1b00a812f95aa9534389b5e42ab155 |
Training suite: libero_90 only. Vision encoder and VLM backbone frozen;
Action Expert + state/action projections trained.
This file
| Field | Value |
|---|---|
run_id |
seen__expert__libero90__nall__s42__20260822T010019Z__gd4b8fb8 |
| step | 100000 |
weights.pt SHA-256 |
2cd510a594a87580f7368b782ca9b37332c0e5002d807093c759e95fbfb57c88 |
| seen-probe success | 24/30 = 0.80 on three frozen libero_90 probes, seeds 1000–1009, horizon 300 |
optimizer.pt |
not uploaded (eval does not need it) |
Selected from libero_90 probes only. Target-task success was not used
to pick this checkpoint, the learning rate, or the training length.
Normalization
Deploy this checkpoint with suite-wide libero_90 MEAN_STD. Do not attach
held-out libero_goal overlay statistics. That mix zeros seen-probe success
and is a deployment incompatibility, not a measure of parameter forgetting.
License
Weights are a derivative of lerobot/smolvla_base trained on
nvidia/LIBERO_LeRobot_v3. Project code is Apache-2.0; follow upstream model
and dataset terms for the weights.