SLIM LIBERO Stage-1 checkpoint
Model weights from the October 12, 2026 reproduction run, after 17,736 optimizer steps (3 epochs) of joint inverse and forward dynamics training.
This is a Stage-1 dynamics checkpoint, intended for Stage-2 policy initialization and representation diagnostics. It is not a final action policy. It is not asserted to be the original Stage-1 ancestor of the paper-reported model or of the existing kzzwang/SLIM-LIBERO release.
Recipe
- Data: LIBERO-90 plus LIBERO-Object, LIBERO-Goal, LIBERO-Spatial, and LIBERO-10.
- Objective: IDM + action-conditioned FDM, loss weights 0.125 : 1.
- Seed: 42; global batch: 128; action horizon: 8.
- Future slots: 512; EMA momentum: 0.999.
- Training source commit:
00426cc82a44ca9397d293fcad226e12dc574d2c. - Code: kzz1031/SLIM.
The included configuration uses environment variables for local assets and data. The original epoch-based recipe resolves to 5,912 updates per epoch and 17,736 updates in total. See provenance.json for checkpoint identity.
Download
hf download kzzwang/SLIM-LIBERO-Stage1 --local-dir ./SLIM-LIBERO-Stage1
sha256sum ./SLIM-LIBERO-Stage1/checkpoints/epoch_3_pytorch_model.pt
Expected SHA-256:
8ec953e7b9bbcacd583f2f7770f8196f0728a4bad1dcabf6acde3b16d4bece3d
The file is a PyTorch model state_dict (1,464,159,013 bytes). It includes the Stage-1 model and EMA target weights, but not optimizer state. For the canonical action-only Stage 2, load the model weights while skipping ema_vision_encoder. and _ema_fp32_ as in the SLIM training implementation, and initialize a new optimizer and schedule.