| --- |
| license: other |
| tags: [lingbot-va, world-model, video-action, robotics, arx-lift-r5, sft-teacher] |
| library_name: diffusers |
| --- |
| |
| # LingBot-VA ARX SFT Teacher — NO train/test split (all 199), step 3000 |
|
|
| Video-action world-model (LingBot-VA) fine-tuned on the **full ARX WAM-OPD |
| dataset (all 199 episodes, no holdout)** — the contract-deliverable teacher. |
|
|
| > This is the **no-split** line. Its training upper bound was **3000 steps** |
| > (contract), so **there is no 5000-step checkpoint for this line** — the |
| > 5000-step run is the separate 160-episode *split* teacher (used only for the |
| > overfitting diagnostic). This checkpoint (step 3000) is the final schedule |
| > checkpoint of the no-split teacher. |
|
|
| ## Provenance |
|
|
| | | | |
| |---|---| |
| | Base model | `Robbyant/lingbot-va-base` @ `5fbd004a…` | |
| | Dataset | `ylhaichen04/WAM-OPD_4Tasks` @ `651338c2…` (LeRobot v2.1, all 199 ep / 64,417 frames) | |
| | Split | none (all 199 for training, no holdout) | |
| | Temporal | `video_7p5hz_action_15hz_k2` (K=2) | |
| | Action | 14-D absolute joint -> 30-D `[14-19,28,21-26,29]`, dataset q01/q99 | |
| | Precision / optim | BF16 FSDP, AdamW lr 5e-6, grad clip 2.0, CFG 0.1 | |
| | This checkpoint | step 3000 (final; no 5000 for this line) | |
|
|
| `transformer/` = diffusers-format WanTransformer3DModel. Switch attn_mode to |
| torch/flashattn for inference (training used flex). |
| |
| ## Selection note |
| |
| Because the no-split line has **no holdout**, step 3000 is a final schedule |
| checkpoint, not a validation-selected best. Checkpoint selection for deployment |
| should be done by real-robot evaluation. |
| |
