neoAI-JP-DeepSeek-Qwen-32B

neoAI-JP-DeepSeek-Qwen-32BはQwen/Qwen2.5-32Bとdeepseek-ai/DeepSeek-R1-Distill-Qwen-32Bをベースにした日本語Reasoningモデルです。 Qwen2.5-32Bに対して日本語で継続事前学習を行ったモデルに対してDeepSeek-R1-Distill-Qwen-32BのChat Vectorをマージして作成しました。

詳細はブログ記事を参照してください。

neoAI 日本語Reasoning Model を開発 Part 1 継続事前学習

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model_name = "neoAI/neoAI-JP-DeepSeek-Qwen-32B"
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
prompt = "How many r's are in the word \"strawberry\""
messages = [
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
print(response)

Developed by

Developers

以下アルファベット順

• Gouki Minegishi
• Kai Yamashita
• Koki Itai
• Masaki Otsuki
• Toshiki Kawamoto

How to Cite

@misc{neoAI-JP-QwQ-32B, 
      title={neoAI-JP-DeepSeek-Qwen-32B}, 
      url={https://huggingface.co/neoai-inc/neoAI-JP-DeepSeek-Qwen-32B}, 
      author={Gouki Minegishi and Kai Yamashita and Koki Itai and Masaki Otsuki and Toshiki Kawamoto},
      year={2025},
}