Modelpublic

hz1919810/lingbot-va-arx-teacher-step3000

sourceHugging Faceotherupdated 26d agoView on Hugging Face

0likes

Model Card

LingBot-VA ARX SFT Teacher — v2 split, step 3000

Video–action world-model (LingBot-VA) fine-tuned on the ARX LIFT R5 dual-arm WAM-OPD task set. This is the transformer of the SFT teacher at optimizer step 3000.

*⚠️ This is the v2 diagnostic run, not the contract deliverable. v2 was trained from base on a 160-episode train split (of 199; per-task 40/10, seed 42) specifically to measure overfitting on held-out data. The contract's full-data teacher is the separate v1 (all-199)* run. Do not conflate the two: use v2 for the overfitting analysis below; use v1 for any all-data delivery.

Provenance


Base model	`Robbyant/lingbot-va-base` @ `5fbd004a61bac17d3764504f09e0b628fc2c0592`
Dataset	`ylhaichen04/WAM-OPD_4Tasks` @ `651338c2…` (LeRobot v2.1, 199 ep / 64,417 frames)
Train subset	160 train episodes (39 held out for validation)
Temporal profile	`video_7p5hz_action_15hz_k2` (K=2, 16 actions/chunk)
Action	14-D absolute joint targets → 30-D layout `[14–19,28,21–26,29]`, q01/q99
Precision / optim	BF16 FSDP, AdamW lr 5e-6, grad clip 2.0, CFG prob 0.1
This checkpoint	step 3000 of a 5000-step schedule

Held-out validation (39 episodes, teacher-forced denoising loss, 4 draws)

step	val latent	val action	val total
1500 (val min)	0.12730	0.005465	0.13276
3000	0.13262	0.005814	0.13843
5000	0.15310	0.006829	0.15993

Overfitting: validation loss bottoms at step 1500 and rises monotonically through 5000 (+~20%). Step 3000 is past the validation minimum (overfit side).

Important caveat on this metric

The validation loss above is teacher-forced, one-step denoising error — the training objective on unseen data. It is an overfitting tripwire, not a task-success ranking. For a world model, low denoising loss does not imply good multi-step rollouts or real-robot success (compounding error is invisible here). Final checkpoint selection requires closed-loop rollout / real-robot evaluation, which for this real-robot dataset is the data owner's call.

Files

transformer/ — diffusers-format WanTransformer3DModel (config.json + bf16 safetensors). Load with the LingBot-VA stack; switch attn_mode to torch/flashattn for inference (training used flex).