jujbob's picture
Update README.md
3185b9c verified
metadata
language:
  - en
  - ko
license: llama3.1
library_name: transformers
base_model:
  - meta-llama/Meta-Llama-3.1-405B

Update!

  • [2024.08.08] preview ๋ชจ๋ธ์ด ์ตœ์ดˆ ์—…๋ฐ์ดํŠธ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค. A100 120๋Œ€ ๊ทœ๋ชจ์˜ ์ปดํ“จํŒ… ํŒŒ์›Œ๋กœ ํ•™์Šต ์ง„ํ–‰์ค‘์œผ๋กœ ๋ชจ๋ธ์€ ๊ณ„์† ์—…๋ฐ์ดํŠธ๋  ์˜ˆ์ •์ž…๋‹ˆ๋‹ค.

Bllossom | Demo | Homepage | Github |

์ €ํฌ Bllossom ํŒ€์—์„œ llama3.1 ๊ธฐ๋ฐ˜์˜ ํ•œ๊ตญ์–ด-์˜์–ด ์ด์ค‘ ์–ธ์–ด๋ชจ๋ธ Bllossom-405B๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค.
์ด๋ฒˆ Bllossom3.1-405B๋Š” preview ๋ฒ„์ „์œผ๋กœ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํŠน์ง•์„ ๋ณด์ž…๋‹ˆ๋‹ค.
 - Llama3.1-405B-Inst ๋Œ€๋น„ 5~10% ํ•œ๊ตญ์–ด ์„ฑ๋Šฅ์ด ํ–ฅ์ƒ ๋˜์—ˆ์Šต๋‹ˆ๋‹ค (single turn ๊ธฐ์ค€).
 - Llama3.1์˜ ์˜์–ด ์„ฑ๋Šฅ์„ ์ „ํ˜€ ์†์ƒ์‹œํ‚ค์ง€ ์•Š์€ ์™„์ „ํ•œ Bilingual ๋ชจ๋ธ์ž…๋‹ˆ๋‹ค.
 - ๊ธฐ์กด ๋ชจ๋ธ ๋Œ€๋น„ ์ž์—ฐ์Šค๋Ÿฝ๊ณ  ์นœ์ ˆํ•œ ํ•œ๊ตญ์–ด ๋ฌธ์žฅ์„ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค.
 - ์ธ๊ฐ„ํ‰๊ฐ€, GPTํ‰๊ฐ€(MT-Bench, LogicKor 9์  ๋“ฑ) ๊ฒฐ๊ณผ GPT4์™€ ์œ ์‚ฌํ•˜๊ฑฐ๋‚˜ ์•ฝ๊ฐ„ ๋‚ฎ์€ ์„ฑ๋Šฅ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.

ํ•ด๋‹น ๋ชจ๋ธ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์€ ํ˜‘์—…์„ ํ† ๋Œ€๋กœ ๊ตฌ์ถ• ๋˜์—ˆ์Šต๋‹ˆ๋‹ค!
 - ์„œ์šธ๊ณผ๊ธฐ๋Œ€ MLP์—ฐ๊ตฌ์‹ค์˜ ๊ฒฝ๋Ÿ‰ํ™” ์‚ฌ์ „ ํ•™์Šต๊ธฐ๋ฒ•์ด ์ ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
 - ํ…Œ๋””์ธ์˜ ์ •๊ตํ•œ Instruction Tuning๊ณผ RAG ๊ธฐ์ˆ ์ด ์ ์šฉ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
 - HP์˜ computing ์ง€์›์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค.
 - Common Crawl ์žฌ๋‹จ์˜ OscarํŒ€์—์„œ ์ ๊ทน์ ์ธ ๋ฐ์ดํ„ฐ ์ง€์›์ด ์žˆ์—ˆ์Šต๋‹ˆ๋‹ค

์–ธ์ œ๋‚˜ ๊ทธ๋žฌ๋“ฏ ํ•ด๋‹น ๋ชจ๋ธ์€ ์ƒ์—…์  ์ด์šฉ์ด ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค. A100 6๋Œ€๋งŒ ์ค€๋น„๋˜๋ฉด Bllossom์„ ์ด์šฉํ•ด ์—ฌ๋Ÿฌ๋ถ„๋งŒ์˜ ๋ชจ๋ธ์„ ๋งŒ๋“ค์–ด๋ณด์„ธ์š” GPT4๊ฐ€ ๋”์ด์ƒ ํ•„์š” ์—†์Šต๋‹ˆ๋‹ค.
GPU์ž์›์ด ๋ถ€์กฑํ•˜๋ฉด A100 3๊ฐœ ํ˜น์€ A6000 4๊ฐœ๋กœ ์–‘์žํ™” ๋ชจ๋ธ์„ ์ด์šฉํ•ด ๋ณด์„ธ์š”. [์–‘์žํ™”๋ชจ๋ธ](https://huggingface.co/MLP-KTLim/llama-3.1-Korean-Bllossom-405B-gguf-Q4_K_M)

1. Bllossom-8B๋Š” ์„œ์šธ๊ณผ๊ธฐ๋Œ€, ํ…Œ๋””์ธ, ์—ฐ์„ธ๋Œ€ ์–ธ์–ด์ž์› ์—ฐ๊ตฌ์‹ค์˜ ์–ธ์–ดํ•™์ž์™€ ํ˜‘์—…ํ•ด ๋งŒ๋“  ์‹ค์šฉ์ฃผ์˜๊ธฐ๋ฐ˜ ๋ฌด๋ฃŒ ์–ธ์–ด๋ชจ๋ธ๋กœ 2023๋…„๋ถ€ํ„ฐ ์ง€์†์ ์ธ ์—…๋ฐ์ดํŠธ๋ฅผ ํ†ตํ•ด ๊ด€๋ฆฌํ•ด ์˜ค๊ณ ์žˆ์Šต๋‹ˆ๋‹ค. ๋งŽ์ด ํ™œ์šฉํ•ด์ฃผ์„ธ์š” ๐Ÿ™‚
2. ์ดˆ ๊ฐ•๋ ฅํ•œ Advanced-Bllossom ๋ชจ๋ธ, ์‹œ๊ฐ-์–ธ์–ด ๋ชจ๋ธ์„ ๋ณด์œ ํ•˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค! (๊ถ๊ธˆํ•˜์‹ ๋ถ„์€ ๊ฐœ๋ณ„ ์—ฐ๋ฝ์ฃผ์„ธ์š”!!)
3. Bllossom์€ NAACL2024, LREC-COLING2024 (๊ตฌ๋‘) ๋ฐœํ‘œ๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
4. ์ข‹์€ ์–ธ์–ด๋ชจ๋ธ ๊ณ„์† ์—…๋ฐ์ดํŠธ ํ•˜๊ฒ ์Šต๋‹ˆ๋‹ค!! ํ•œ๊ตญ์–ด ๊ฐ•ํ™”๋ฅผ์œ„ํ•ด ๊ณต๋™ ์—ฐ๊ตฌํ•˜์‹ค๋ถ„(ํŠนํžˆ๋…ผ๋ฌธ) ์–ธ์ œ๋“  ํ™˜์˜ํ•ฉ๋‹ˆ๋‹ค!! 
   ๊ทธ๋ฆฌ๊ณ  ์†Œ๋Ÿ‰์˜ GPU๋ผ๋„ ๋Œ€์—ฌ ๊ฐ€๋Šฅํ•œํŒ€์€ ์–ธ์ œ๋“  ์—ฐ๋ฝ์ฃผ์„ธ์š”! ๋งŒ๋“ค๊ณ  ์‹ถ์€๊ฑฐ ๋„์™€๋“œ๋ ค์š”.
The Bllossom language model is a Korean-English bilingual language model based on the open-source LLama3.1. It enhances the connection of knowledge between Korean and English. It has the following features:
 - Korean performance improved by 5-10% compared to Llama 3.1-405B-Inst (on Single Turn Eval).
 - A complete bilingual model that does not compromise the English performance of Llama 3.1.
 - Generates more natural and friendly Korean sentences compared to existing models.
 - Human evaluations and GPT evaluations (MT-Bench, LogicKor scoring 9, etc.) show performance similar to or slightly lower than GPT-4.

This model developed by MLPLab at Seoultech, Teddysum and Yonsei Univ

Example code

Colab Tutorial

Install Dependencies

pip install torch transformers==4.40.0 accelerate

Python code with Pipeline

import transformers
import torch

model_id = "Bllossom/llama-3.1-Korean-Bllossom-405B"

pipeline = transformers.pipeline(
    "text-generation",
    model=model_id,
    model_kwargs={"torch_dtype": torch.bfloat16},
    device_map="auto",
)

pipeline.model.eval()

PROMPT = '''You are a helpful AI assistant. Please answer the user's questions kindly. ๋‹น์‹ ์€ ์œ ๋Šฅํ•œ AI ์–ด์‹œ์Šคํ„ดํŠธ ์ž…๋‹ˆ๋‹ค. ์‚ฌ์šฉ์ž์˜ ์งˆ๋ฌธ์— ๋Œ€ํ•ด ์นœ์ ˆํ•˜๊ฒŒ ๋‹ต๋ณ€ํ•ด์ฃผ์„ธ์š”.'''
instruction = "์„œ์šธ์˜ ์œ ๋ช…ํ•œ ๊ด€๊ด‘ ์ฝ”์Šค๋ฅผ ๋งŒ๋“ค์–ด์ค„๋ž˜?"

messages = [
    {"role": "system", "content": f"{PROMPT}"},
    {"role": "user", "content": f"{instruction}"}
    ]

prompt = pipeline.tokenizer.apply_chat_template(
        messages, 
        tokenize=False, 
        add_generation_prompt=True
)

terminators = [
    pipeline.tokenizer.eos_token_id,
    pipeline.tokenizer.convert_tokens_to_ids("<|eot_id|>")
]

outputs = pipeline(
    prompt,
    max_new_tokens=2048,
    eos_token_id=terminators,
    do_sample=True,
    temperature=0.6,
    top_p=0.9
)

print(outputs[0]["generated_text"][len(prompt):])
# ๋ฌผ๋ก ์ด์ฃ ! ์„œ์šธ์€ ๋‹ค์–‘ํ•œ ๋ฌธํ™”์™€ ์—ญ์‚ฌ, ์ž์—ฐ์„ ๊ฒธ๋น„ํ•œ ๋„์‹œ๋กœ, ๋งŽ์€ ๊ด€๊ด‘ ๋ช…์†Œ๋ฅผ ์ž๋ž‘ํ•ฉ๋‹ˆ๋‹ค. ์—ฌ๊ธฐ ์„œ์šธ์˜ ์œ ๋ช…ํ•œ ๊ด€๊ด‘ ์ฝ”์Šค๋ฅผ ์†Œ๊ฐœํ•ด ๋“œ๋ฆด๊ฒŒ์š”.

### ์ฝ”์Šค 1: ์—ญ์‚ฌ์™€ ๋ฌธํ™” ํƒ๋ฐฉ

1. **๊ฒฝ๋ณต๊ถ**
   - ์„œ์šธ์˜ ๋Œ€ํ‘œ์ ์ธ ๊ถ๊ถ๋กœ, ์กฐ์„  ์™•์กฐ์˜ ์—ญ์‚ฌ์™€ ๋ฌธํ™”๋ฅผ ์ฒดํ—˜ํ•  ์ˆ˜ ์žˆ๋Š” ๊ณณ์ž…๋‹ˆ๋‹ค.

2. **๋ถ์ดŒ ํ•œ์˜ฅ๋งˆ์„**
   - ์ „ํ†ต ํ•œ์˜ฅ์ด ์ž˜ ๋ณด์กด๋œ ๋งˆ์„๋กœ, ์กฐ์„ ์‹œ๋Œ€์˜ ์ƒํ™œ์ƒ์„ ๋Š๋‚„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

...

Supported by

  • Hewlett Packard (HP) Enterprise
  • Common Crawl
  • AICA

Citation

Language Model

@misc{bllossom,
  author = {ChangSu Choi, Yongbin Jeong, Seoyoon Park, InHo Won, HyeonSeok Lim, SangMin Kim, Yejee Kang, Chanhyuk Yoon, Jaewan Park, Yiseul Lee, HyeJin Lee, Younggyun Hahm, Hansaem Kim, KyungTae Lim},
  title = {Optimizing Language Augmentation for Multilingual Large Language Models: A Case Study on Korean},
  year = {2024},
  journal = {LREC-COLING 2024},
  paperLink = {\url{https://arxiv.org/pdf/2403.10882}},
 },
}

Vision-Language Model

@misc{bllossom-V,
  author = {Dongjae Shin, Hyunseok Lim, Inho Won, Changsu Choi, Minjun Kim, Seungwoo Song, Hangyeol Yoo, Sangmin Kim, Kyungtae Lim},
  title = {X-LLaVA: Optimizing Bilingual Large Vision-Language Alignment},
  year = {2024},
  publisher = {GitHub},
  journal = {NAACL 2024 findings},
  paperLink = {\url{https://arxiv.org/pdf/2403.11399}},
 },
}

Contact

Contributor