jangmin's picture
Update README.md
88ead7d
---
license: apache-2.0
library_name: peft
---
# Food Order Understanding in Korean
This is a LoRA adapter as a result of fine-tuning the pre-trained model `meta-llama/Llama-2-13b-chat-hf`. It is designed with the expectation of understanding Korean food ordering sentences, and analyzing food menus, option names, and quantities.
## Usage
Here is an example of loading the model.
Note the pretrained model is `meta-llama/Llama-2-13b-chat-hf`.
```python
peft_model_id = "jangmin/qlora-llama2-13b-chat-hf-food-order-understanding-30K"
config = PeftConfig.from_pretrained(peft_model_id)
bnb_config = BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_use_double_quant=True,
bnb_4bit_quant_type="nf4",
bnb_4bit_compute_dtype=torch.bfloat16
)
model = AutoModelForCausalLM.from_pretrained(config.base_model_name_or_path, quantization_config=bnb_config, cache_dir=cache_dir, device_map={"":0})
model = PeftModel.from_pretrained(model, peft_model_id)
tokenizer = AutoTokenizer.from_pretrained(config.base_model_name_or_path, cache_dir=cache_dir)
model.eval()
```
Inferece can be done as follows.
```python
instruction_prompt_template = """
๋‹ค์Œ์€ ๋งค์žฅ์—์„œ ๊ณ ๊ฐ์ด ์Œ์‹์„ ์ฃผ๋ฌธํ•˜๋Š” ์ฃผ๋ฌธ ๋ฌธ์žฅ์ด๋‹ค. ์ด๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์Œ์‹๋ช…, ์˜ต์…˜, ์ˆ˜๋Ÿ‰์„ ์ถ”์ถœํ•˜์—ฌ ๊ณ ๊ฐ์˜ ์˜๋„๋ฅผ ์ดํ•ดํ•˜๊ณ ์ž ํ•œ๋‹ค.
๋ถ„์„ ๊ฒฐ๊ณผ๋ฅผ ์™„์„ฑํ•ด์ฃผ๊ธฐ ๋ฐ”๋ž€๋‹ค.
### ์ฃผ๋ฌธ ๋ฌธ์žฅ: {0} ### ๋ถ„์„ ๊ฒฐ๊ณผ:
"""
def gen(x):
q = instruction_prompt_template.format(x)
gened = model.generate(
**tokenizer(
q,
return_tensors='pt',
return_token_type_ids=False
).to('cuda'),
max_new_tokens=256,
early_stopping=True,
do_sample=True,
eos_token_id=tokenizer.eos_token_id
)
decoded_results = tokenizer.batch_decode(gened, skip_special_tokens=True)
return decoded_results[0]
```
A generated sample is as follows.
```python
print(gen("์•„์ด์Šค์•„๋ฉ”๋ฆฌ์นด๋…ธ ํ†จ์‚ฌ์ด์ฆˆ ํ•œ์ž” ํ•˜๊ณ ์š”. ๋”ธ๊ธฐ์Šค๋ฌด๋”” ํ•œ์ž” ์ฃผ์„ธ์š”. ๋˜, ์ฝœ๋“œ๋ธŒ๋ฃจ๋ผ๋–ผ ํ•˜๋‚˜์š”."))
```
```
๋‹ค์Œ์€ ๋งค์žฅ์—์„œ ๊ณ ๊ฐ์ด ์Œ์‹์„ ์ฃผ๋ฌธํ•˜๋Š” ์ฃผ๋ฌธ ๋ฌธ์žฅ์ด๋‹ค. ์ด๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์Œ์‹๋ช…, ์˜ต์…˜๋ช…, ์ˆ˜๋Ÿ‰์„ ์ถ”์ถœํ•˜์—ฌ ๊ณ ๊ฐ์˜ ์˜๋„๋ฅผ ์ดํ•ดํ•˜๊ณ ์ž ํ•œ๋‹ค.
๋ถ„์„ ๊ฒฐ๊ณผ๋ฅผ ์™„์„ฑํ•ด์ฃผ๊ธฐ ๋ฐ”๋ž€๋‹ค.
### ๋ช…๋ น: ์•„์ด์Šค์•„๋ฉ”๋ฆฌ์นด๋…ธ ํ†จ์‚ฌ์ด์ฆˆ ํ•œ์ž” ํ•˜๊ณ ์š”. ๋”ธ๊ธฐ์Šค๋ฌด๋”” ํ•œ์ž” ์ฃผ์„ธ์š”. ๋˜, ์ฝœ๋“œ๋ธŒ๋ฃจ๋ผ๋–ผ ํ•˜๋‚˜์š”. ### ์‘๋‹ต:
- ๋ถ„์„ ๊ฒฐ๊ณผ 0: ์Œ์‹๋ช…:์•„์ด์Šค์•„๋ฉ”๋ฆฌ์นด๋…ธ, ์˜ต์…˜:ํ†จ์‚ฌ์ด์ฆˆ, ์ˆ˜๋Ÿ‰:ํ•œ์ž”
- ๋ถ„์„ ๊ฒฐ๊ณผ 1: ์Œ์‹๋ช…:๋”ธ๊ธฐ์Šค๋ฌด๋””, ์ˆ˜๋Ÿ‰:ํ•œ์ž”
- ๋ถ„์„ ๊ฒฐ๊ณผ 2: ์Œ์‹๋ช…:์ฝœ๋“œ๋ธŒ๋ฃจ๋ผ๋–ผ, ์ˆ˜๋Ÿ‰:ํ•˜๋‚˜
``````
## Training
Fine-tuning was conducted using https://github.com/artidoro/qlora on an RTX-4090 machine, and took approximately 9 hours.
The max_steps parameter was set to 5,000, which allowed nearly two complete scans of the entire dataset.
Below is my training script.
```bash
python qlora.py \
--cache_dir /Jupyter/huggingface/.cache \
--model_name_or_path meta-llama/Llama-2-13b-chat-hf \
--use_auth \
--output_dir ../output/llama2-gpt4-30k-food-order-understanding-13b \
--logging_steps 10 \
--save_strategy steps \
--data_seed 42 \
--save_steps 500 \
--save_total_limit 40 \
--evaluation_strategy steps \
--eval_dataset_size 1024 \
--max_eval_samples 1000 \
--per_device_eval_batch_size 12 \
--max_new_tokens 32 \
--dataloader_num_workers 1 \
--group_by_length \
--logging_strategy steps \
--remove_unused_columns False \
--do_train \
--do_eval \
--lora_r 64 \
--lora_alpha 16 \
--lora_modules all \
--double_quant \
--quant_type nf4 \
--bf16 \
--bits 4 \
--warmup_ratio 0.03 \
--lr_scheduler_type constant \
--gradient_checkpointing \
--dataset /Jupyter/dev_src/ASR-for-noisy-edge-devices/data/food-order-understanding-gpt4-30k.json \
--target_max_len 512 \
--per_device_train_batch_size 12 \
--gradient_accumulation_steps 1 \
--max_steps 5000 \
--eval_steps 500 \
--learning_rate 0.0002 \
--adam_beta2 0.999 \
--max_grad_norm 0.3 \
--lora_dropout 0.1 \
--weight_decay 0.0 \
--seed 0 \
--report_to tensorboard
```
## Dataset
The dataset was constructed using GPT-API with `gpt-4`. A prompt template is desginged to generate examples of sentence pairs of a food order and its understanding. Total 30k examples were generated.
Note that it cost about $400 to generate 30K examples through 3,000 API calls.
Some generated examples are as follows:
```json
{
'input': '๋‹ค์Œ์€ ๋งค์žฅ์—์„œ ๊ณ ๊ฐ์ด ์Œ์‹์„ ์ฃผ๋ฌธํ•˜๋Š” ์ฃผ๋ฌธ ๋ฌธ์žฅ์ด๋‹ค. ์ด๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์Œ์‹๋ช…, ์˜ต์…˜๋ช…, ์ˆ˜๋Ÿ‰์„ ์ถ”์ถœํ•˜์—ฌ ๊ณ ๊ฐ์˜ ์˜๋„๋ฅผ ์ดํ•ดํ•˜๊ณ ์ž ํ•œ๋‹ค.\n๋ถ„์„ ๊ฒฐ๊ณผ๋ฅผ ์™„์„ฑํ•ด์ฃผ๊ธฐ ๋ฐ”๋ž€๋‹ค.\n\n### ๋ช…๋ น: ์ œ์œก๋ณถ์Œ ํ•œ๊ทธ๋ฆ‡ํ•˜๊ณ ์š”, ๋น„๋น”๋ฐฅ ํ•œ๊ทธ๋ฆ‡ ์ถ”๊ฐ€ํ•ด์ฃผ์„ธ์š”. ### ์‘๋‹ต:\n',
'output': '- ๋ถ„์„ ๊ฒฐ๊ณผ 0: ์Œ์‹๋ช…:์ œ์œก๋ณถ์Œ,์ˆ˜๋Ÿ‰:ํ•œ๊ทธ๋ฆ‡\n- ๋ถ„์„ ๊ฒฐ๊ณผ 1: ์Œ์‹๋ช…:๋น„๋น”๋ฐฅ,์ˆ˜๋Ÿ‰:ํ•œ๊ทธ๋ฆ‡'
},
{
'input': '๋‹ค์Œ์€ ๋งค์žฅ์—์„œ ๊ณ ๊ฐ์ด ์Œ์‹์„ ์ฃผ๋ฌธํ•˜๋Š” ์ฃผ๋ฌธ ๋ฌธ์žฅ์ด๋‹ค. ์ด๋ฅผ ๋ถ„์„ํ•˜์—ฌ ์Œ์‹๋ช…, ์˜ต์…˜๋ช…, ์ˆ˜๋Ÿ‰์„ ์ถ”์ถœํ•˜์—ฌ ๊ณ ๊ฐ์˜ ์˜๋„๋ฅผ ์ดํ•ดํ•˜๊ณ ์ž ํ•œ๋‹ค.\n๋ถ„์„ ๊ฒฐ๊ณผ๋ฅผ ์™„์„ฑํ•ด์ฃผ๊ธฐ ๋ฐ”๋ž€๋‹ค.\n\n### ๋ช…๋ น: ์‚ฌ์ฒœํƒ•์ˆ˜์œก ๊ณฑ๋ฐฐ๊ธฐ ์ฃผ๋ฌธํ•˜๊ณ ์š”, ์ƒค์›Œํฌ๋ฆผ์น˜ํ‚จ๋„ ํ•˜๋‚˜ ์ถ”๊ฐ€ํ•ด์ฃผ์„ธ์š”. ### ์‘๋‹ต:\n',
'output': '- ๋ถ„์„ ๊ฒฐ๊ณผ 0: ์Œ์‹๋ช…:์‚ฌ์ฒœํƒ•์ˆ˜์œก,์˜ต์…˜:๊ณฑ๋ฐฐ๊ธฐ\n- ๋ถ„์„ ๊ฒฐ๊ณผ 1: ์Œ์‹๋ช…:์ƒค์›Œํฌ๋ฆผ์น˜ํ‚จ,์ˆ˜๋Ÿ‰:ํ•˜๋‚˜'
}
```
## Note
I have another fine-tuned Language Model, `jangmin/qlora-polyglot-ko-12.8b-food-order-understanding-32K`, which is based on `EleutherAI/polyglot-ko-12.8b`. The dataset was generated using `gpt-3.5-turbo-16k`. I believe that the quality of a dataset generated by `GPT-4` would be superior to that generated by `GPT-3.5`.