cagataydev/smolvla_tictactoe_vision_unfrozen

SmolVLA TicTacToe -- Vision Unfrozen (60K steps)

Fine-tuned SmolVLA with vision encoder unfrozen (393M trainable params) on TicTacToe manipulation data.

Training Recipe

• Base: lerobot/smolvla_base
• Steps: 60,000 / 60,000
• Trainable params: 392,904,096 (393M) -- full model including vision
• Samples seen: 480,000 (~3.32 epochs)
• Duration: 640.4 min (~10h40m)
• Final loss: 0.078
• Final gradient norm: 2.175
• LR (final): 2.5e-06

Load

Notes

Vision encoder was not frozen during fine-tuning, allowing full-model adaptation. Compare with frozen-vision baseline for ablation.