LingBot-VLA 2.0 — LIBERO full fine-tune, step 1000
Full fine-tune (nothing frozen) continued from the OpenRoboto LIBERO checkpoint. Bittensor subnet 80 (OpenRoboto), simulation competition 2.
Provenance
| base | openroboto-ai/lingbot-vla-v2-6b-libero @ ce6a322157acc7a03d0ca71bb84423c7f2e124d7 (its step 6000) |
| training code | github.com/Robbyant/lingbot-vla-v2 @ 951475ae1b1d87553e7dc47c97b53a3d695c0d13 |
| dataset | lerobot/libero @ a1aaacb7f6cd6ee5fb43120f673cebb0cfea7dd4 (1693 episodes / 273465 frames) |
| processor | Qwen/Qwen3-VL-4B-Instruct @ ebb281ec70b05090aa6165b016eac8ec08e71b17 |
| config | the vendor's own training/libero_full.yaml, shipped inside the base checkpoint |
Training
2x NVIDIA H100 PCIe 80GB, FSDP2 full-shard, bf16 with fp32 master weights.
- batch: micro 8 x grad-accum 4 x 2 GPUs = global 64 (matches the vendor's
train_args.sh, which used micro 8 x accum 1 x 8 GPUs) - optimizer muon, lr 5e-5 cosine -> 5e-6, warmup 2%
max_steps: 1500set deliberately so the cosine anneals fully — this is a self-contained 1500-step run, not a truncated 30000-step schedule- gradient checkpointing on, MoE (32 experts, top-4), chunk_size 50
- depth + video alignment losses active (MoGe + MDM + DINO-video teachers)
- 12.55 s/step, 5 h 13 m wall clock for 1500 steps
Training loss 0.1544 (step 1) -> ~0.085. Metrics: https://wandb.ai/atomyuri46-weights-biases/sn80-lingbot-fullft
Weight drift vs the base
Relative L2 over all 1708 tensors:
| checkpoint | overall | trunk model.qwenvl_with_expert |
|---|---|---|
| step 500 | 2.431 % | 2.443 % |
| step 1000 | 2.880 % | 2.889 % |
| step 1500 | 2.928 % | 2.934 % |
Drift saturates: 500->1000 adds 0.45 pp, 1000->1500 only 0.05 pp. step 1000 and step 1500 are near-duplicates.
Status
Passes openroboto check (24337.8 MB, within the 20-35 GB gate).
Not yet evaluated on LIBERO — no local scoring harness existed at upload time,
so the success rate of this checkpoint is unknown, and a falling training loss is
not evidence of a better policy.
- Downloads last month
- 8