Bi-LaWM RoboTwin Stage3 Finetune — From Stage2-Only Ablation 50k
Repository: GT-111/bi-lawm-robotwin-finetune-from-stage2only (renamed from bi-lawm-robotwin-stage3-finetune on 2026-08-31)
Intermediate checkpoints from the RoboTwin Stage3 finetune training run of Bi-LaWM (80k-step plan, stopped at 50k by decision).
Lineage
- Initialized from the consolidated Stage2 50k export of the stage2-only ablation (
export/step-050000/pytorch_model.pt). - Milestones are published every 5k steps plus a snapshot of the latest checkpoint.
Checkpoints
| Hub path | Training step | Format |
|---|---|---|
checkpoints/step-005000/ |
5,000 | Full sharded FSDP training checkpoint |
checkpoints/step-010000/ |
10,000 | Full sharded FSDP training checkpoint |
checkpoints/step-015000/ |
15,000 | Full sharded FSDP training checkpoint |
checkpoints/step-020000/ |
20,000 | Full sharded FSDP training checkpoint |
checkpoints/step-025000/ |
25,000 | Full sharded FSDP training checkpoint |
checkpoints/step-030000/ |
30,000 | Full sharded FSDP training checkpoint |
checkpoints/step-035000/ |
35,000 | Full sharded FSDP training checkpoint |
checkpoints/step-040000/ |
40,000 | Full sharded FSDP training checkpoint |
checkpoints/step-045000/ |
45,000 | Full sharded FSDP training checkpoint |
checkpoints/step-050000/ |
50,000 | Final checkpoint (80k plan stopped at 50k by decision) |
Each checkpoint contains the model shards, per-rank optimizer and RNG states, scheduler state, and trainer_state.json. These are resumable distributed-training checkpoints rather than merged inference-only weights.
Experiment
- Project: Bi-LaWM
- Dataset/environment: RoboTwin
- Stage: Stage3 policy finetune
- Policy initialization: Stage2 50k ablation export (see
GT-111/bi-lawm-stage2-only-ablation) - Training plan: 80k steps (stopped at 50k by decision)
Loading
Use these checkpoints with the matching Bi-LaWM code and FSDP configuration. For inference or evaluation, export/merge the selected distributed checkpoint with the project checkpoint tooling first.