| Standford Alpaca | |
| FINETUNED USING THE ORIGINAL REPOSITORY: https://github.com/tatsu-lab/stanford_alpaca | |
| NO LORA HAS BEEN USED | |
| status of training can be viewed at: https://wandb.ai/peruano/huggingface/runs/ei57qbzm | |
| SCRIPT TO CONVERT: https://gist.github.com/eous/31959971768a0b56a5fdb1c7db85c6e3 | |
| CONFIGURATION (default): | |
| ```shell | |
| torchrun --nproc_per_node=4 --master_port=3045 train.py \ | |
| --model_name_or_path /workspace/llama-7b-hf \ | |
| --data_path ./alpaca_data.json \ | |
| --bf16 True \ | |
| --output_dir /workspace/output \ | |
| --num_train_epochs 3 \ | |
| --per_device_train_batch_size 4 \ | |
| --per_device_eval_batch_size 4 \ | |
| --gradient_accumulation_steps 8 \ | |
| --evaluation_strategy "no" \ | |
| --save_strategy "steps" \ | |
| --save_steps 200 \ | |
| --save_total_limit 1 \ | |
| --learning_rate 2e-5 \ | |
| --weight_decay 0. \ | |
| --warmup_ratio 0.03 \ | |
| --lr_scheduler_type "cosine" \ | |
| --logging_steps 1 \ | |
| --fsdp "full_shard auto_wrap" \ | |
| --fsdp_transformer_layer_cls_to_wrap 'LLaMADecoderLayer' \ | |
| --tf32 True --report_to="wandb" | |
| ``` |