Downloads · 30 days
0
alexsuw/smolvla-libero-fewshot-naive-baseline
smolvla-libero-fewshot-naive-baseline is a robotics model from alexsuw. 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 other.
Thirty independent naive target fine-tunes of the frozen seen-expert checkpoint. Each cell starts from the same origin weights and sees only a prefix of one frozen liberogoal demonstration list. N=10 does not continue…
Downloads · 30 days
0
Access
Public
Updated Aug 23, 2026
Repo size
27.2 GB
Likes
0
Public
Click a slice to open those files.
.pt27.2 GB · 100%
From the Hugging Face model README
Thirty independent naive target fine-tunes of the frozen seen-expert
checkpoint. Each cell starts from the same origin weights and sees only a
prefix of one frozen libero_goal demonstration list. N=10 does not
continue N=5.
Code: github.com/alexsuw/smolvla-libero-fewshot
Origin seen checkpoint:
alexsuw/smolvla-libero-fewshot-seen-expert-100k
(weights.pt SHA-256 2cd510a594a87580f7368b782ca9b37332c0e5002d807093c759e95fbfb57c88)
Collection:
alexsuw/smolvla-libero-few-shot-6a8b009357482d2b4b9d3c2f
Held-out LIBERO-Goal instructions (exact NVIDIA LeRobot v3 text):
| slug | instruction | dataset task_index |
|---|---|---|
drawer_middle | open the middle drawer of the cabinet | 9 |
bowl_stove | put the bowl on the stove | 7 |
wine_cabinet | put the wine bottle on top of the cabinet | 4 |
nvidia/LIBERO_LeRobot_v3 revision e5907374380b8f96511957e6ba5582be52a1e179libero_goalconfigs/splits/target_splits.json (first 1 / 2 / 5 / 10 / 25){42, 123}lerobot/smolvla_base revision c83c3163b8ca9b7e67c509fffd9121e66cb96205<task>_n<NN>_s<seed>/
weights.pt
normalization_stats.json # subset overlay MEAN_STD used at train/eval
COMPLETED.json
checksums.json
config.resolved.yaml
trainable_parameters.txt
Example:
from huggingface_hub import hf_hub_download
hf_hub_download(
"alexsuw/smolvla-libero-fewshot-naive-baseline",
"bowl_stove_n10_s42/weights.pt",
)
hf_hub_download(
"alexsuw/smolvla-libero-fewshot-naive-baseline",
"bowl_stove_n10_s42/normalization_stats.json",
)
optimizer.pt is not uploaded. index.json lists Hub prefixes, training
steps, and weights.pt SHA-256 for every cell.
Pooled over 3 tasks × 2 seeds. N=0 is zero-shot from the seen checkpoint
(not in this repo). Official spec budgets are {5,10,25}; {1,2} use the
same prefixes and the same warmup_steps=1000 (those short runs stay inside
warmup — a recorded limitation, not a retune).
| N | naive target success |
|---|---|
| 0 | 1/60 (0.017) |
| 1 | 109/120 (0.908) |
| 2 | 100/120 (0.833) |
| 5 | 107/120 (0.892) |
| 10 | 116/120 (0.967) |
| 25 | 114/120 (0.950) |
Target success was not used to choose hyperparameters or the seen checkpoint.
normalization_stats.json (train overlay).libero_90 suite stats with these
adapted weights if you want a weight-forgetting comparison.Derivatives of lerobot/smolvla_base trained on nvidia/LIBERO_LeRobot_v3.
Project code is Apache-2.0; follow upstream model and dataset terms for weights.