Reliable VLA Fine-Tuning & Evaluation

4-GPU SmolVLA training and RNG-controlled LIBERO evaluation

I built an end-to-end SmolVLA pipeline that validates the native LIBERO data contract, trains with 4-GPU DDP, evaluates through the official closed-loop protocol, and makes stochastic checkpoint comparisons attributable.

Impact

  • 3.82× DDP speedup on 4×RTX 4090 at matched per-GPU batch
  • 1,693 episodes / 273,465 frames / 40 LIBERO tasks
  • Identified a flow-matching RNG confound that changed the apparent direction of checkpoint scaling
  • Rebuilt the comparison with identical environment starts and per-episode policy-noise streams

The controlled 5K→20K result was 2/30→4/30. I report this as a weak positive signal rather than a high-performing-policy claim; the main contribution is a training and evaluation workflow whose conclusions can be trusted.

Code and full evidence