hf_dataset/: serialized Hugging Face dataset used by GRPO notebook
sft_splits/: train/val/test split ids used for SFT
grpo_splits/: train/val/test split ids used for GRPO
image_cache_448px.tar.zst: cached 448px resized images for training
To restore the cache locally:
tar --zstd -xf image_cache_448px.tar.zst
