metadata
license: apache-2.0
pipeline_tag: text2text-generation
tags:
- llama
- llm
This is LoRA checkpoint fine-tuned with the following CLI. The fine-tuning process is logged in W&B dashboard. I have used DGX workstation with 8 x A100(40G).
python finetune.py \
--base_model='elinas/llama-65b-hf-transformers-4.29' \
--data_path='alpaca_data.json' \
--num_epochs=10 \
--cutoff_len=1024 \
--group_by_length \
--output_dir='./lora-alpaca-65b-elinas' \
--lora_target_modules='[q_proj,k_proj,v_proj,o_proj]' \
--lora_r=16 \
--lora_alpha=32 \
--batch_size=1024 \
--micro_batch_size=15
This LoRA checkpoint is recommended to be used with transformers >= 4.29 which should be installed with the following command currently(4/30/2023).
pip install git+https://github.com/huggingface/transformers.git