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.
- Downloads last month
- -