Finetuned LoRA model with command $llmtune finetune --model llama-30b-4bit --weights llama-30b-4bit.pt --dataset data50.json --adapter alpaca-adapter-folder-30b-4bit
Using first 50 records of Alpaca dataset from original dataset.json The training loss is almost flat, and result is clearly better than original llama-30b-4bit model.
License shall follow typical Alpaca family models situation.