SmolVLA N=1 Frozen-Stats FT and Anchored FT (L2-SP)

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
Collection: SmolVLA LIBERO Few-shot

Frozen protocol and results

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.

Layout and loading

<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.

License

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.