harpreet22happy/granite-3.3-8b-cuad-lora

cuad long-context SFT LoRA

LoRA adapter for ibm-granite/granite-3.3-8b-instruct, trained on the cuad long-document task at up to 24576 tokens per example (fine_tune_experiments/task3/ in the source repo: govreport = report -> summary, kleister = charity report -> 8-field JSON, cuad = contract -> 41-category clause JSON). main = best checkpoint by held-out loss; branch final = last. Stage: SFT.

• LoRA: r=16, alpha=32, target_modules=['q_proj', 'k_proj', 'v_proj', 'o_proj', 'gate_proj', 'up_proj', 'down_proj']
• bf16, FlashAttention-2 + padding-free packing, completion-only loss (report masked, summary supervised)

This branch: checkpoint-50 (dev loss 0.0442, the minimum of 5 evals).

Dev loss by step: 10: 0.0632, 20: 0.0520, 30: 0.0470, 40: 0.0446, 50: 0.0442