train-sft

name: train-sft description: Run supervised fine-tuning with veRL

Run supervised fine-tuning for protein-LLM:

1. Verify environment:

   source /home/yeopjin/orcd/pool/init_protein_llm.sh
   python -c "import torch, verl, flash_attn; print(f'GPUs: {torch.cuda.device_count()}')"
   

2. Check configuration:

   • Read configs/sft_config.yaml (if exists)
   • Verify LoRA settings: r=8, applied to k/v matrices
   • Confirm ESM-3 is frozen

3. Pre-flight checks:

   • GPU memory: nvidia-smi
   • Data exists: ls ./data/pdb_2021aug02_sample/
   • Wandb configured: echo $WANDB_DIR

4. Run training:

   python scripts/train_sft.py --config configs/sft_config.yaml
   

5. Monitor:

   • Check wandb dashboard for loss curves
   • Watch for OOM errors
   • Verify gradient norms are stable

6. Post-training validation:

   • Run evaluation on validation set
   • Check checkpoint saved: ls outputs/checkpoints/