metadata

library_name: peft
license: apache-2.0
tags:
  - llama2
  - qLoRa
  - traditional_chinese
  - alpaca
  - adapter

Traditional Chinese Llama2

github repo: https://github.com/MIBlue119/traditional_chinese_llama2/
Practice to finetune Llama2 on traditional chinese instruction dataset at Llama2 chat model. I use qlora and the alpaca translated dataset to finetune llama2-7b model at rtx3090(24GB VRAM) with 9 hours.

Thanks for these references:

NTU NLP Lab's alapaca dataset: alpaca-tw_en-align.json: ntunpllab translate Stanford Alpaca 52k dataset
Chinese Llama 2 7B train.py
Load the pretrained model in 4-bit precision and Set training with LoRA according to hf's trl lib: QLoRA finetuning

Resources

traditional chinese qlora finetuned Llama2 merge model: weiren119/traditional_chinese_qlora_llama2_merged
traditional chinese qlora adapter model: weiren119/traditional_chinese_qlora_llama2

Online Demo

Run the qlora finetuned model at colab: May need colab pro or colab pro+

Notice

the repois model adpater if you want to use the merged checkpoint(adapter+original model) repo: https://huggingface.co./weiren119/traditional_chinese_qlora_llama2_merged

Use which pretrained model

NousResearch: https://huggingface.co./NousResearch/Llama-2-7b-chat-hf

Training procedure

The following bitsandbytes quantization config was used during training:

load_in_8bit: False
load_in_4bit: True
llm_int8_threshold: 6.0
llm_int8_skip_modules: None
llm_int8_enable_fp32_cpu_offload: False
llm_int8_has_fp16_weight: False
bnb_4bit_quant_type: nf4
bnb_4bit_use_double_quant: True
bnb_4bit_compute_dtype: bfloat16

Framework versions

PEFT 0.4.0

Usage

Installation dependencies

$pip install transformers torch peft

Run the inference

import transformers
import torch
from transformers import AutoTokenizer, TextStreamer
from peft import AutoPeftModelForCausalLM

# Use the same tokenizer from the source model
original_model_path="NousResearch/Llama-2-7b-chat-hf"
tokenizer = AutoTokenizer.from_pretrained(original_model_path, use_fast=False)

# Load qlora fine-tuned model, you can replace this with your own model
qlora_model_path = "weiren119/traditional_chinese_qlora_llama2"

model = AutoPeftModelForCausalLM.from_pretrained(
        qlora_model_path,
        load_in_4bit=qlora_model_path.endswith("4bit"),
        torch_dtype=torch.float16,
        device_map='auto'
    )

system_prompt = """You are a helpful, respectful and honest assistant. Always answer as helpfully as possible, while being safe.  Your answers should not include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal content. Please ensure that your responses are socially unbiased and positive in nature.

            If a question does not make any sense, or is not factually coherent, explain why instead of answering something not correct. If you don't know the answer to a question, please don't share false information."""




def get_prompt(message: str, chat_history: list[tuple[str, str]]) -> str:
    texts = [f'[INST] <<SYS>>\n{system_prompt}\n<</SYS>>\n\n']
    for user_input, response in chat_history:
        texts.append(f'{user_input.strip()} [/INST] {response.strip()} </s><s> [INST] ')
    texts.append(f'{message.strip()} [/INST]')
    return ''.join(texts)


print ("="*100)
print ("-"*80)
print ("Have a try!")

s = ''
chat_history = []
while True:
    s = input("User: ")
    if s != '':
        prompt = get_prompt(s, chat_history)
        print ('Answer:')
        tokens = tokenizer(prompt, return_tensors='pt').input_ids
        #generate_ids = model.generate(tokens.cuda(), max_new_tokens=4096, streamer=streamer)
        generate_ids = model.generate(input_ids=tokens.cuda(), max_new_tokens=4096, streamer=streamer)
        output = tokenizer.decode(generate_ids[0, len(tokens[0]):-1]).strip()
        chat_history.append([s, output])
        print ('-'*80)