Downloads · 30 days
0
alexsuw/smolvla-libero-fewshot-stability-n1
smolvla-libero-fewshot-stability-n1 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.
Twelve independently trained checkpoints from a matched stability experiment: two methods, three held-out LIBERO-Goal tasks, and train seeds 42/123. All cells start from the same frozen seen-expert-100k checkpoint and…
Downloads · 30 days
0
Access
Public
Updated Aug 23, 2026
Repo size
10.9 GB
Likes
0
Public
Click a slice to open those files.
.pt10.9 GB · 100%
From the Hugging Face model README
Twelve independently trained checkpoints from a matched stability experiment:
two methods, three held-out LIBERO-Goal tasks, and train seeds 42/123. All
cells start from the same frozen
seen-expert-100k
checkpoint and use the first registered target demonstration.
Code: github.com/alexsuw/smolvla-libero-fewshot<br> Collection: SmolVLA LIBERO Few-shot
Both methods use the original libero_90 normalization during target training
and deployment. No target-overlay statistics are fitted. Anchored FT adds an
FP32 raw-sum L2-SP penalty with preregistered lambda=0.01 on the trainable
Action Expert/projection parameters relative to their frozen initialization.
| Method | Target success | Corrected seen retention | Trainable parameters | Peak VRAM |
|---|---|---|---|---|
| Naive N=1 reference | 109/120 (90.8%) | 37/180 (20.6%) | 99,880,992 | 7,540 MiB |
| Frozen-Stats FT N=1 | 109/120 (90.8%) | 39/180 (21.7%) | 99,880,992 | 7,540 MiB |
| Anchored FT N=1 | 105/120 (87.5%) | 57/180 (31.7%) | 99,880,992 | 8,036 MiB |
No method or anchoring strength was selected, tuned, or rerun using target or retention success.
<frozen_stats|anchored_l2sp>/<task>_n01_s<seed>/
weights.pt
normalization_stats.json # canonical frozen libero_90 statistics
config.resolved.yaml
COMPLETED.json
checksums.json
trainable_parameters.txt
run/ # provenance and training metrics, no optimizer
from huggingface_hub import hf_hub_download
weights = hf_hub_download(
"alexsuw/smolvla-libero-fewshot-stability-n1",
"anchored_l2sp/drawer_middle_n01_s42/weights.pt",
)
stats = hf_hub_download(
"alexsuw/smolvla-libero-fewshot-stability-n1",
"anchored_l2sp/drawer_middle_n01_s42/normalization_stats.json",
)
Every published normalization file has SHA-256
b159b6fed3e52edf25bd39b377dd64940221b7a030362daf7f726b1c2ecb30cf.
index.json records exact checkpoint hashes and per-cell evaluation results.
Full tables and limitations are in results/results.md.
Optimizer/RNG state, raw rollouts, datasets, traces, videos, and credentials are intentionally excluded.
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.