brian-lim's picture
Add How to use
20c3afe
metadata
license: apache-2.0
datasets:
  - brian-lim/smile_style_orca
language:
  - ko

Korean Style Transfer

This model is a fine-tuned version of Synatra-7B-v0.3-dpo using a Korean style dataset provided by Smilegate AI (https://github.com/smilegate-ai/korean_smile_style_dataset/tree/main). Since the original dataset is tabular and not fit for training the LLM, I have preprocessed it into an instruction-input-output format, which can be found here. The dataset is then fed into the ChatML template. Feel free to use my version of the dataset as needed.

ν•΄λ‹Ή λͺ¨λΈμ€ Synatra-7B-v0.3-dpo λͺ¨λΈμ„ 슀마일게이트 AIμ—μ„œ μ œκ³΅ν•˜λŠ” Smile style λ°μ΄ν„°μ…‹μœΌλ‘œ νŒŒμΈνŠœλ‹ ν–ˆμŠ΅λ‹ˆλ‹€. κΈ°μ‘΄ 데이터셋은 ν…Œμ΄λΈ” ν˜•νƒœλ‘œ λ˜μ–΄μžˆμ–΄ ν•΄λ‹Ή 데이터λ₯Ό instruction-input-output ν˜•νƒœλ‘œ λ§Œλ“€μ—ˆκ³ , μ—¬κΈ°μ—μ„œ 확인 κ°€λŠ₯ν•©λ‹ˆλ‹€. 데이터셋을 뢈러온 λ’€ ChatML ν˜•μ‹μ— 맞좰 ν›ˆλ ¨ 데이터 ꡬ좕을 ν•œ λ’€ μ§„ν–‰ν–ˆμŠ΅λ‹ˆλ‹€. ν•„μš”ν•˜μ‹œλ‹€λ©΄ 자유둭게 μ‚¬μš©ν•˜μ‹œκΈ° λ°”λžλ‹ˆλ‹€.

How to use

>>> import torch
>>> from transformers import AutoModelForCausalLM, AutoTokenizer

device = 'cuda' if torch.cuda.is_available() else 'cpu'

tokenizer = AutoTokenizer.from_pretrained('brian-lim/smile-style-transfer')
model = AutoModelForCausalLM.from_pretrained('brian-lim/smile-style-transfer', device_map = device)

prompts = {'informal': '주어진 글을 κ°€λŠ₯ν•œ ν˜•μ‹μ μ΄μ§€ μ•Šκ³  λ”±λ”±ν•˜μ§€ μ•Šμ€ λŒ€ν™”μ²΄λ‘œ λ°”κΏ”μ€˜.',
          'android': '주어진 글을 κ°€λŠ₯ν•œ μ•ˆλ“œλ‘œμ΄λ“œ λ‘œλ΄‡κ³Ό 같은 λŒ€ν™”μ²΄λ‘œ λ°”κΏ”μ€˜.',
          'azae': '주어진 글을 κ°€λŠ₯ν•œ 아저씨같은 말투둜 λ°”κΏ”μ€˜.',
          'chat': '주어진 글을 κ°€λŠ₯ν•œ 인터넷상에 μ‚¬μš©ν•˜λŠ” 말투둜 λ°”κΏ”μ€˜.',
          'choding': '주어진 글을 κ°€λŠ₯ν•œ μ΄ˆλ“±ν•™μƒμ²˜λŸΌ 짧게 쀄인 λŒ€ν™”μ²΄λ‘œ λ°”κΏ”μ€˜.',
          'emoticon': '주어진 글을 κ°€λŠ₯ν•œ 이λͺ¨ν‹°μ½˜μ΄ λ“€μ–΄κ°„ λŒ€ν™”μ²΄λ‘œ λ°”κΏ”μ€˜.',
          'enfp': '주어진 글을 κ°€λŠ₯ν•œ ν™œκΈ°μ°¨λ©΄μ„œ 곡감을 많이 ν•˜λŠ” μΉœμ ˆν•œ λŒ€ν™”μ²΄λ‘œ λ°”κΏ”μ€˜.',
          'gentle' : '주어진 글을 κ°€λŠ₯ν•œ β€œμš”β€λ‘œ λλ‚˜μ§€ μ•ŠμœΌλ©΄μ„œ κΉ”λ”ν•œ λŒ€ν™”μ²΄λ‘œ λ°”κΏ”μ€˜.',
          'halbae' : '주어진 글을 κ°€λŠ₯ν•œ μ—°λ₯œμ΄ μžˆλŠ” 할아버지 같은 맑투둜 λ°”κΏ”μ€˜.',
          'halmae' : '주어진 글을 κ°€λŠ₯ν•œ 비속어가 λ“€μ–΄κ°€λŠ” ν• λ¨Έλ‹ˆ 같은 맑투둜 λ°”κΏ”μ€˜.',
          'joongding': '주어진 글을 κ°€λŠ₯ν•œ 쀑학ꡐ 2ν•™λ…„μ˜ 말투둜 λ°”κΏ”μ€˜.',
          'king': '주어진 글을 κ°€λŠ₯ν•œ μ‘°μ„ μ‹œλŒ€ μ™•μ˜ 말투둜 λ°”κΏ”μ€˜.',
          'seonbi': '주어진 글을 κ°€λŠ₯ν•œ μ‘°μ„ μ‹œλŒ€ μ„ λΉ„μ˜ 말투둜 λ°”κΏ”μ€˜.',
          'sosim': '주어진 글을 κ°€λŠ₯ν•œ μ•„μ£Ό μ†Œμ‹¬ν•˜κ³  μ‘°μ‹¬μŠ€λŸ¬μš΄ 말투둜 λ°”κΏ”μ€˜.',
          'translator': '주어진 글을 κ°€λŠ₯ν•œ μ–΄μƒ‰ν•œ ν•œκ΅­μ–΄ λ²ˆμ—­ 말투둜 λ°”κΏ”μ€˜.',
          }
query = '[INPUT]: μ•ˆλ…•ν•˜μ„Έμš”. μš”μ¦˜ 날씨가 많이 μŒ€μŒ€ν•˜λ„€μš” \n[OUTPUT]: '

input_query = prompts['king'] + query
input_tokenized = tokenizer(input_query,return_tensors="pt").to(device)

g_config = GenerationConfig(temperature=0.3,
                        repetition_penalty=1.2,
                        max_new_tokens=768,
                        do_sample=True,
                        )
output = model.generate(**input_tokenized,
                      generation_config=g_config,        
                      pad_token_id=tokenizer.eos_token_id,
                      eos_token_id=tokenizer.eos_token_id,)
output_text = tokenizer.decode(output.detach().cpu().numpy()[0])
output_text = output_text[output_text.find('[OUTPUT]'):]
print(output_text)

license: apache-2.0