Model Card for RedWhale-tv-10.8B-v1.0

Model Description

RedWhale은 전처리한 한국어 Corpus, 특화된 한국어 Tokenizer, 효과적인 Model initialization, Continuous Multi-Stage Pretraining strategy 등을 갖추고 있습니다. 이러한 접근 방식은 높은 정확도와 이해도를 유지하면서 Computational costs를 줄여 제한된 리소스에서 Pretraining을 가능하게 해줍니다. RedWhale 사용을 원하시면 repo access 요청해주세요.

About the Model

• Name: TwinDoc/RedWhale-tv-10.8B-v1.0
• Foundation Model: upstage/SOLAR-10.7B-v1.0
• Train Corpus: preprocessed AI-Hub datasets
• Developed by: 애자일소다 (AGILESODA)
• Model type: llama
• Language(s) (NLP): 한국어, 영어
• License: cc-by-nc-sa-4.0
• Paper: RedWhale: An Adapted Korean LLM Through Efficient Continual Pretraining

Load the Model

from transformers import AutoTokenizer
from transformers import AutoModelForCausalLM

YOUR_HF_TOKEN_READ = "hf_..."
model_name_or_path = "TwinDoc/RedWhale-tv-10.8B-v1.0"
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, token=YOUR_HF_TOKEN_READ)
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, token=YOUR_HF_TOKEN_READ)

Generate Text

text = "대한민국의 수도는"
encodings = tokenizer(text, return_tensors='pt')
terminators = [tokenizer.eos_token_id] + tokenizer("\n", add_special_tokens=False)["input_ids"]

outputs = model.generate(**encodings, eos_token_id=terminators)
generated_text = tokenizer.batch_decode(outputs)[0]
# '<s> 대한민국의 수도는 서울이다.\n'

License

The content of this project, created by AGILESODA, is licensed under the Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0).

Citation

@misc{vo2024redwhaleadaptedkoreanllm,
      title={RedWhale: An Adapted Korean LLM Through Efficient Continual Pretraining}, 
      author={Anh-Dung Vo and Minseong Jung and Wonbeen Lee and Daewoo Choi},
      year={2024},
      eprint={2408.11294},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2408.11294}, 
}

Built with: