Modelpublic

lygwm-review/LY-GWM-RoboCasa-Human300

sourceHugging Faceupdated 3d agoView on Hugging Face

0likes

Model Card

LY-GWM-RoboCasa-Human300 — S42 weights

Public inference weights for the S42 four-candidate learned-reward evaluation. Code and technical report: https://github.com/SEIN-LYGWM/LY-GWM-RoboCasa-Human300-eval

S42: 207/2500 (8.28%), RoboCasa 1.0.1, split pretrain. Atomic seen 150/900 (16.6667%); composite seen 38/800 (4.750%); composite unseen 19/800 (2.375%).

Path	Component
`gr00t/`	R5 fine-tuned GR00T-N1.5 checkpoint-120000; retained existing files
`lygwm/best.pt`	W5 action-conditioned graph dynamics; retained existing file
`s42/scorer-best.pt`	S38 binary state-reward head used for S42 selection
`s42/scorer-training-config.json`	Original scorer configuration pinned by S42 inference

GR00T generates four candidates; LY-GWM predicts their future features/state; S38 scores them; the highest reward logit selects the action chunk. S38 has no explicit goal input and is not a calibrated long-horizon success model. Use reviewer_s42/ in the code repository; this is not an automatic Transformers pipeline. Pin the HF revision containing the S42 additions when reproducing. Tensor SHA256 values are in release_info.json.

The original cluster run completed; the new portable entrypoint has static/CPU checks only, without a completed GPU reproduction on a second machine. Compiled environments and simulator assets are separate dependencies. Full historical raw action traces are not included in the code publication subset.

Preserve existing component notices and weight terms. Public availability does not create a replacement license; see MODEL_PROVENANCE.md. This update adds inference artifacts, not a complete training reproduction release.