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

Task and data

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

Dataset: nvidia/LIBERO_LeRobot_v3 revision e5907374380b8f96511957e6ba5582be52a1e179
Suite for these runs: libero_goal
Episode IDs: nested prefixes of configs/splits/target_splits.json (first 1 / 2 / 5 / 10 / 25)
Train seeds: {42, 123}
Base policy: lerobot/smolvla_base revision c83c3163b8ca9b7e67c509fffd9121e66cb96205
Papers: SmolVLA, LIBERO

Layout

text

<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:

python

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.

Target success (20 eval seeds / 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 (read this)

Target eval: use this folder’s normalization_stats.json (train overlay).
Seen-probe / retention: use libero_90 suite stats with these adapted weights if you want a weight-forgetting comparison.
Evaluating adapted weights on seen probes with the target overlay is a deployment mismatch. That number is not catastrophic forgetting.

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.