doanthuan / bloom-gpt

Instruction Tuning Large Language Model for Vietnamese

Bloom-GPT

Instruction Tuning Large Language Model for Vietnamese

The general architecture and experimental results of our method can be found in our paper:

@InProceedings{,
title     = "{Efficient Finetuning Large Language Models For Vietnamese Chatbot}",
author    = {Vu-Thuan Doan, Quoc-Truong Truong, Duc-Vu Nguyen, Vinh-Tiep Nguyen and Thuy-Ngan Nguyen Luu},
booktitle = {In Proceedings of International Conference on Multimedia Analysis and Pattern Recognition (MAPR)},
year      = {2023},
pages     = "{to appear}"
}

Setup Enviroment

1. Install dependencies

   pip install -r requirements.txt

2. If bitsandbytes doesn't work, install it from source. Windows users can follow these instructions.

Download Dataset

• instruct_merged.jsonl: instruction dataset. It contains 52k samples from Alpaca + 170k samples from GPT4All. Then translated to Vietnamese.

  wget https://storage.googleapis.com/doanthuan/data/instruct_merged.jsonl

• translated_health_200k.jsonl: Medical instruction dataset. It was collected from ChatDoctor

  wget https://storage.googleapis.com/doanthuan/data/translated_health_200k.jsonl

Training

Finetune Bloomz-Chat:

python finetune_bloomz_instruct.py \
    --base_model 'bigscience/bloomz-7b1-mt' \
    --data_path 'instruct_merged.jsonl' \
    --output_dir './bloomz-instruct'

Finetune Bloomz-Doctor:

python finetune_bloomz_doctor.py \
    --base_model 'bigscience/bloomz-7b1-mt' \
    --data_path 'translated_health_200k.jsonl' \
    --output_dir './bloomz-doctor'

We can also tweak our hyperparameters:

python finetune_bloomz_instruct.py \
    --base_model 'bigscience/bloomz-7b1-mt' \
    --data_path 'instruct_merged.jsonl' \
    --output_dir './bloomz-instruct' \
    --batch_size 128 \
    --micro_batch_size 4 \
    --num_epochs 3 \
    --learning_rate 1e-4 \
    --cutoff_len 512 \
    --val_set_size 2000 \
    --lora_r 8 \
    --lora_alpha 16 \
    --lora_dropout 0.05 \
    --lora_target_modules '[q_proj,v_proj]' \
    --train_on_inputs \
    --group_by_length

Inference (demo)

Bloom-Chat demo

Bloom-Doctor demo

In demo directory also contains notebooks for a demo.

Model weights

LoRA weights are in models directory