grpo_stable_reasoning_0709
GRPO RL checkpoints for a SimLingo-style driving VLA (InternVL2-1B backbone + continuous
speed-waypoint head). Experiment: grpo_stable_reasoning (0709 run).
Each stepNNN/pytorch_model.pt is a consolidated fp32 state_dict (converted from the
DeepSpeed ZeRO training checkpoint at that global step).
| file | step |
|---|---|
step100/pytorch_model.pt |
100 |
step200/pytorch_model.pt |
200 |
step300/pytorch_model.pt |
300 |
Load with:
import torch
sd = torch.load("step300/pytorch_model.pt", map_location="cpu")