GR00T real-world multitask QCRL (fail500_ps)
Artifacts for failure_reward=-500 with primitive_success shaping on the last step
(failure_progress_penalty_scale=500).
| Stage | Content |
|---|---|
| returns | meta/*/returns_groot_aiha_multitask_fail500_ps.parquet |
| value SFT | value/global_step_18000/... |
| advantages | meta/*/advantages_groot_aiha_multitask_fail500_ps_N16_ckpt18000_q30.parquet |
| CFG SFT (QAT) | cfg/global_step_*/... |
Base policy: sm-park/aiha_multitask_groot (GR00T N1.5).
Load value checkpoint
export GROOT_RW_VALUE_CKPT="/path/to/this-repo/value/global_step_18000/actor/model_state_dict"
Load CFG policy
export GROOT_RW_MODEL_PATH="/path/to/this-repo/cfg/global_step_<STEP>/actor"
# RLinf resolves actor/model_state_dict/full_weights.pt under model_path.
Dataset sidecars
Upload matching meta/*.parquet into:
sm-park/multitask_combined_v3(SFT)sm-park/groot_multitask_FP_rolloutsm-park/groot_multitask_PTQ_rollout
See manifest.csv for exact paths and sizes.