Twelve independently trained checkpoints: two methods, three held-out target tasks, and train seeds 42/123. Every cell starts from the same frozen seen-expert-100k checkpoint (SHA-256 2cd510a594a87580f7368b782ca9b37332c0e5002d807093c759e95fbfb57c88) and uses only the first registered target demonstration.

Code: github.com/alexsuw/smolvla-libero-fewshot
Collection: SmolVLA LIBERO Few-shot

Methods and results

Method	Target success	Corrected seen retention	Trainable parameters
Naive N=1 reference	109/120 (90.8%)	37/180 (20.6%)	99,880,992
Target-LoRA N=1	99/120 (82.5%)	19/180 (10.6%)	4,215,632
Replay-LoRA N=1	67/120 (55.8%)	2/180 (1.1%)	4,215,632

Replay-LoRA uses a fixed 75% target / 25% seen mixture. Seen replay comes only from libero_90 and was not selected using the three retention probes. No method was tuned or rerun using success.

Layout and loading

text

<target_lora|replay_lora>/<task>_n01_s<seed>/
  weights.pt
  adapter/adapter_model.pt
  adapter/adapter_config.json
  normalization_stats.json
  config.resolved.yaml
  COMPLETED.json
  checksums.json
  trainable_parameters.txt
  run/                         # provenance and training metrics, no optimizer

weights.pt is the complete inference checkpoint. The adapter/ files are also provided for adapter-specific analysis. index.json contains exact weights/adapter hashes and per-cell target/retention results.

python

from huggingface_hub import hf_hub_download

weights = hf_hub_download(
    "alexsuw/smolvla-libero-fewshot-lora-n1",
    "target_lora/drawer_middle_n01_s42/weights.pt",
)
adapter = hf_hub_download(
    "alexsuw/smolvla-libero-fewshot-lora-n1",
    "target_lora/drawer_middle_n01_s42/adapter/adapter_model.pt",
)

For target deployment use the cell's normalization_stats.json. Corrected seen retention must instead use the frozen libero_90 statistics with digest b159b6fed3e52edf25bd39b377dd64940221b7a030362daf7f726b1c2ecb30cf.

Optimizer/RNG state, raw rollouts, datasets, traces, videos, and credentials are intentionally excluded. Full tables and limitations are in results/results.md.

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.