Update!
- [2024.10.08] Bllossom-3B λͺ¨λΈμ΄ μ΅μ΄ μ λ°μ΄νΈ λμμ΅λλ€.
Bllossom | Demo | Homepage | Github |
μ ν¬ Bllossom νμμ Bllossom-3B λͺ¨λΈμ 곡κ°ν©λλ€.
llama3.2-3Bκ° λμλλ° νκ΅μ΄κ° ν¬ν¨ μλμλ€κ΅¬?? μ΄λ² Bllossom-3Bλ νκ΅μ΄κ° μ§μλμ§ μλ κΈ°λ³Έ λͺ¨λΈμ νκ΅μ΄-μμ΄λ‘ κ°νλͺ¨λΈμ
λλ€.
- 100% full-tuningμΌλ‘ 150GBμ μ μ λ νκ΅μ΄λ‘ μΆκ° μ¬μ νμ΅ λμμ΅λλ€. (GPUλ§μ΄ νμ μ΅λλ€)
- κ΅μ₯ν μ μ λ Instruction Tuningμ μ§ννμ΅λλ€.
- μμ΄ μ±λ₯μ μ ν μμμν€μ§ μμ μμ ν Bilingual λͺ¨λΈμ
λλ€.
- LogicKor κΈ°μ€ 5Bμ΄ν μ΅κ³ μ μλ₯Ό κΈ°λ‘νκ³ 6μ μ΄λ°λ μ μλ₯Ό 보μ
λλ€.
- Instruction tuningλ§ μ§ννμ΅λλ€. DPO λ± μ±λ₯ μ¬λ¦΄ λ°©λ²μΌλ‘ νλν΄λ³΄μΈμ.
- MT-Bench, LogicKor λ± λ²€μΉλ§ν¬ μ μλ₯Ό μλ°κΈ° μν΄ μ λ΅λ°μ΄ν°λ₯Ό νμ©νκ±°λ νΉμ λ²€μΉλ§ν¬λ₯Ό νκ²ν
ν΄μ νμ΅νμ§ μμμ΅λλ€. (ν΄λΉ λ²€μΉλ§ν¬ νκ²ν
ν΄μ νμ΅νλ©΄ 8μ λ λμ΅λλ€...)
μΈμ λ κ·Έλ¬λ― ν΄λΉ λͺ¨λΈμ μμ
μ μ΄μ©μ΄ κ°λ₯ν©λλ€.
1. Bllossomμ AAAI2024, NAACL2024, LREC-COLING2024 (ꡬλ) λ°νλμμ΅λλ€.
2. μ’μ μΈμ΄λͺ¨λΈ κ³μ μ
λ°μ΄νΈ νκ² μ΅λλ€!! νκ΅μ΄ κ°νλ₯Όμν΄ κ³΅λ μ°κ΅¬νμ€λΆ(νΉνλ
Όλ¬Έ) μΈμ λ νμν©λλ€!!
from llama_cpp import Llama
from transformers import AutoTokenizer
model_id = 'Bllossom/llama-3.2-Korean-Bllossom-3B'
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = Llama(
model_path='llama-3.2-Korean-Bllossom-3B-gguf-Q4_K_M.gguf'
)
instruction = "μ² μκ° 20κ°μ μ°νμ κ°μ§κ³ μμλλ° μν¬κ° μ λ°μ κ°μ Έκ°κ³ λ―Όμκ° λ¨μ 5κ°λ₯Ό κ°μ Έκ°μΌλ©΄ μ² μμκ² λ¨μ μ°νμ κ°―μλ λͺκ°μΈκ°μ?"
messages = [
{"role": "user", "content": f"{instruction}"}
]
prompt = tokenizer.apply_chat_template(
messages,
tokenize = False,
add_generation_prompt=True
)
generation_kwargs = {
"max_tokens":512,
"stop":["<|eot_id|>"],
"echo":True,
"top_p":0.9,
"temperature":0.6,
}
resonse_msg = model(prompt, **generation_kwargs)
print(resonse_msg['choices'][0]['text'][len(prompt):])
μ² μκ° 20κ°μ μ°νμ κ°μ§κ³ μμκ³ μν¬κ° μ λ°μ κ°μ Έκ°λ©΄, μν¬κ° κ°μ Έκ° μ°νμ κ°―μλ 20 / 2 = 10κ°μ
λλ€.
μ΄μ μ² μκ° λ¨μ μ°νμ κ°―μλ₯Ό κ³μ°ν΄λ³΄κ² μ΅λλ€. μν¬κ° 10κ°λ₯Ό κ°μ Έκ° ν μ² μκ° λ¨μ μ°νμ κ°―μλ 20 - 10 = 10κ°μ
λλ€.
λ―Όμκ° λ¨μ 5κ°λ₯Ό κ°μ Έκ°μΌλ―λ‘, μ² μκ° λ¨μ μ°νμ κ°―μλ 10 - 5 = 5κ°μ
λλ€.
λ°λΌμ μ² μκ° λ¨μ μ°νμ κ°―μλ 5κ°μ
λλ€.
Supported by
- 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
- μκ²½ν(KyungTae Lim), Professor at Seoultech.
ktlim@seoultech.ac.kr
- ν¨μκ· (Younggyun Hahm), CEO of Teddysum.
hahmyg@teddysum.ai
- κΉνμ(Hansaem Kim), Professor at Yonsei.
khss@yonsei.ac.kr
Contributor
- μ νκ²°(Hangyeol Yoo), hgyoo@seoultech.ac.kr
- μ λμ¬(Dongjae Shin), dylan1998@seoultech.ac.kr
- μνμ(Hyeonseok Lim), gustjrantk@seoultech.ac.kr
- μμΈνΈ(Inho Won), wih1226@seoultech.ac.kr
- κΉλ―Όμ€(Minjun Kim), mjkmain@seoultech.ac.kr
- μ‘μΉμ°(Seungwoo Song), sswoo@seoultech.ac.kr
- μ‘μ ν(Jeonghun Yuk), usually670@gmail.com
- μ΅μ°½μ(Chansu Choi), choics2623@seoultech.ac.kr
- μ‘μν(Seohyun Song), alexalex225225@gmail.com
- Downloads last month
- 637