| license: mit | |
| Alpaca Lora adapter weight fine-tuned on following instruction dataset. | |
| https://huggingface.co/datasets/rewoo/planner_instruction_tuning_2k/blob/main/README.md | |
| Training script: borrowed from the official [Alpaca-LoRA](https://github.com/tloen/alpaca-lora) implementation | |
| We use following parameter. | |
| ``` | |
| python finetune.py \ | |
| --base_model 'decapoda-research/llama-7b-hf' \ | |
| --data_path 'rewoo/planner_instruction_tuning_2k' \ | |
| --output_dir './lora-alpaca-planner' \ | |
| --batch_size 128 \ | |
| --micro_batch_size 8 \ | |
| --num_epochs 10 \ | |
| --learning_rate 1e-4 \ | |
| --cutoff_len 1024 \ | |
| --val_set_size 200 \ | |
| --lora_r 8 \ | |
| --lora_alpha 16 \ | |
| --lora_dropout 0.05 \ | |
| --lora_target_modules '[q_proj,v_proj]' \ | |
| --train_on_inputs \ | |
| --group_by_length \ | |
| --resume_from_checkpoint 'tloen/alpaca-lora-7b' | |
| ``` |