Modelpublic

tomoniyukiwo/smolvla-libero-checkpoint-comparison

sourceHugging Faceapache-2.0updated 11d agoView on Hugging Face

0likes34downloads

Model Card

tomoniyukiwo/smolvla-libero-checkpoint-comparison

SmolVLAをLIBEROでfine-tuningし、学習進行に伴うpolicyの変化を比較する公開checkpoint repoです。

Comparison video

Panel	Step	Source
Initial	0	`lerobot/smolvla_base`
1/3 steps	33,333	`checkpoints/033333/pretrained_model`
2/3 steps	66,666	`checkpoints/066666/pretrained_model`
Final	100,000	`checkpoints/100000/pretrained_model`

Evaluation rollout: libero_spatial, task id 0, seed 1000, episode cap 300 steps.

The four panels use the same task, seed, initial-state behavior, and episode cap. Shorter videos are padded with their last frame only for side-by-side playback.

Intended use

This repository is intended for checkpoint progression analysis and reproducible LIBERO evaluation. The single rollout shown in the video is qualitative; use multiple episodes and seeds for quantitative claims.

Training

—Base: lerobot/smolvla_base
—Dataset: lerobot/libero
—Total loop steps: 100,000
—Batch size: 1
—Gradient accumulation: 8
—Precision: BF16
—Vision encoder: frozen
—Empty camera placeholders: 0
—Image rename: image → camera1, image2 → camera2

Evaluation JSON and the four source videos are stored below artifacts/.