ABEJA QwQ32b Reasoning Japanese V1.0

🚀 ABEJA-QwQ32b-Reasoning-Japanese-v1.0

ABEJA-QwQ32b-Reasoning-Japanese-v1.0 is a reasoning model based on abeja/ABEJA-Qwen2.5-32b-Japanese-v0.1().*

By merging the ChatVector of Qwen/QwQ-32B into ABEJA-Qwen2.5-32b-Japanese-v0.1 and then performing additional training, the model ensures Japanese language performance as a reasoning model.

(*A model that has undergone continuous pre-training centered on the Japanese language based on Qwen/Qwen2.5-32B-Instruct.)

It provides the final output after the thinking process enclosed by <think></think>.

🚀 Quick Start

✨ Features

ABEJA-QwQ32b-Reasoning-Japanese-v1.0 inherits the characteristics of QwQ-32B used in model merging. It is recommended to follow the Usage-Guideline of QwQ-32B.

https://huggingface.co/Qwen/QwQ-32B#usage-guidelines

⚠️ Important Note

• Please start the output after <think>\n to enforce the thinking process. If you use apply_chat_template with add_generation_prompt=True, it will be automatically applied.
• It is recommended to use parameters such as Temperature=0.6, TopP=0.95, MinP=0, and TopK between 20 and 40. (We have confirmed that changing these values significantly will reduce accuracy.)
• In multi-turn conversations, only include the final output in the conversation history and do not include the thinking process enclosed by <think></think>. This feature is already included in apply_chat_template.
• A system prompt is not required. Please start with role:user.

💻 Usage Examples

Basic Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "abeja/ABEJA-QwQ32b-Reasoning-Japanese-v1.0"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "人とAIが協調するためには？"
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 = model.generate(
    **model_inputs,
    max_new_tokens=32768,
    do_sample=True,
    temperature=0.6,
    top_k=40,
    top_p=0.95,
)
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)

📄 License

The project is licensed under the Apache-2.0 license.

👨‍💻 Developers

• Hiroshi Kiyota
• Keisuke Fujimoto
• Kentaro Nakanishi
• Kyo Hattori
• Shinya Otani
• Shogo Muranushi
• Takuma Kume
• Tomoki Manabe

(*) In alphabetical order

📋 Model Information