grpo_stable_reasoning_nolow_0722_075_1200steps
GRPO RL checkpoint for a SimLingo-style driving VLA (InternVL2-1B + continuous speed-waypoint head). Variant: nolow, ratio 0.75, 1200-step schedule (baseline temperature).
checkpoints/step1000.ckpt/pytorch_model.pt= consolidated fp32state_dict(step 1000)..hydra/config.yaml= training config — required to rebuild the model for eval.
import torch
sd = torch.load("checkpoints/step1000.ckpt/pytorch_model.pt", map_location="cpu")
Only step 1000 is published; other checkpoints and
rollouts.jsonlare not included.