lingbotvla-rlinf-sft

lingbotvla-rlinf-sft

Backup of the LingBot-VLA 4B RoboTwin place_shoe SFT run (lingbotvla_sft_official_like_7gpu_gb252_clean_eval500_20260718_224152, RLinf robotwin_sft_lingbotvla config).

Contents

• base_model/lingbot-vla-4b/ — LingBot-VLA 4B base weights (official release, 16G), used as actor.model.model_path / rollout model.
• checkpoints/global_step_27500/ — best checkpoint, clean eval success 22.9%. Contains actor/model_state_dict/full_weights.pt only (eval / cold-start fine-tune; no optimizer state).
• checkpoints/global_step_52000/ — latest complete checkpoint, clean eval success 8.3%. Contains both actor/model_state_dict/full_weights.pt and actor/dcp_checkpoint/ (FSDP shards + .metadata, resumable with optimizer state).
• tensorboard/ — full TensorBoard event files of the run (train + post-save eval scalars).
• tensorboard_backup_before_purge_20260728_223851/ — earlier TensorBoard backup before a purge.
• metadata/ — retention/eval records (checkpoint_retention.tsv, best_success_once.tsv), launch and post-save-eval scripts, resume command reference.

Note: global_step_52500 was a truncated save (missing .metadata, training interrupted mid-write) and is intentionally excluded.

Training setup (for resume)

• Base model: base_model/lingbot-vla-4b/ (this repo) + Qwen2.5-VL-3B-Instruct processor
• Dataset: robotwin_lerobot_v21_novideo_place_shoe_clean50 (LeRobot v2.1, no video, 50 clean episodes)
• Norm stats: lingbot-vla/assets/norm_stats/robotwin_all_new.json
• Code: RLinf fork (RLinf_lingbotvla_sft_fix) + lingbot-vla; eval env: RoboTwin sim assets
• See metadata/launch_sft_gb252_clean_eval500.sh for the exact launch command.