v2 domain SFT: early-stopped at step 1100 (val_loss 1.886), trained on the expanded 40k-example dataset -- see README for honest limitations (not a confirmed improvement over v1 on hard questions)

v1 domain SFT: grounded QA finetune (PM-KISAN/schemes/banking domain), early-stopped at step 200 -- see README for known limitations

Upload README.md with huggingface_hub

Upload folder using huggingface_hub

initial commit