grpo_stable_reasoning_05_nnear10_nolow_0710
GRPO RL checkpoints for a SimLingo-style driving VLA (InternVL2-1B + continuous speed-waypoint head).
- Training script:
run_grpo_stable_reasoning_05_nolow.sh - Each
checkpoints/stepNNN.ckpt/pytorch_model.pt= consolidated fp32state_dict. .hydra/config.yaml= the training config (needed to rebuild the model for eval).
| file | step |
|---|---|
checkpoints/step100.ckpt/pytorch_model.pt | 100 |
checkpoints/step150.ckpt/pytorch_model.pt | 150 |
checkpoints/step200.ckpt/pytorch_model.pt | 200 |
checkpoints/step250.ckpt/pytorch_model.pt | 250 |
checkpoints/step300.ckpt/pytorch_model.pt | 300 |
checkpoints/step400.ckpt/pytorch_model.pt | 400 |
checkpoints/step450.ckpt/pytorch_model.pt | 450 |
import torch
sd = torch.load("checkpoints/step300.ckpt/pytorch_model.pt", map_location="cpu")