This repository is based on tloen/alpaca-lora.
instruct_tuning.py is used for tuning.
WANDB_PROJECT=<PROOJECT_NAME> WANDB_RUN_NAME=<RUN_NAME> \
poetry run python -m torch.distributed.launch --nproc_per_node=2 --node_rank=0 train.py \
--data_path izumi-lab/llm-japanese-dataset \
--output_dir <directory_to_save_checkpoints> \
--run_name llama7b \
--parameter_file params.json \
--local_rank=0
fastchat-cli.py is used for chat in CLI.
poetry run python fastchat-cli.py \
--model-path <path_to_model_directory> \
(--lora-weight <path_to_lora_directory>) \
(--temperature 0.7) # default: 0 \
(--conv-template japanese)
For LoRA model, adapter_config.json must be in the directory of the model's weight (checkpoint).
Use --conv-template dolly_v2
with MPT-7B instruct.
For more detail about conv-template, please see FastChat/fastchat/conversation.py
If you would like to chat with rinna/jaapnese-gpt-neox-3.6b-instruction-sft:
poetry run python -m jallm.models.rinna_instruct_sft.cli
# detault temperature value is 0.01 (cannot set to 0.0)
Only no-context mode is available (cannot inherit old prompts).
adapter_config.json must be placed (or copied) in ~/model_weights/lora_weight.
poetry run python -m jallm.utils.convert_pytorch_to_adapter \
--model ~/model_weights/llama-7b \
--lora-weight ~/model_weights/lora_weight \
--output-dir ~/model_weights/converted_weights
poetry run python evaluate.py \
--task {ppl-vqa, jnli} \
--model-path decapoda-research/llama-13b-hf \
--lora-weight lora-weights \
--device cuda \
--format-lang ja \
--max-length 256
@preprint{Suzuki2023-jainsttuning,
title={{From Base to Conversational: Japanese Instruction Dataset and Tuning Large Language Models}},
author={Masahiro Suzuki and Masanori Hirano and Hiroki Sakaji},
year={2023},
doi={10.48550/arXiv.2309.03412},
arxivId={2309.03412},
archivePrefix={arXiv},
}
The codes in this repository are distributed under MIT.