Qwen2.5-3B-Instruct-ko

Introduction

This model is an adaptation of the Qwen2.5-3B model fine-tuned for the Korean language. Compared to Qwen2.5-3B-Instruct, it demonstrates improved performance on Korean benchmarks, leveraging techniques such as SFT, Model Merging, and iterative online DPO. You can easily test it using the code provided below. Special thanks to Alibaba for consistently providing excellent models!

Quickstart

Here provides a code snippet with apply_chat_template to show you how to load the tokenizer and model and how to generate contents.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "tryumanshow/Qwen2.5-3B-Instruct-ko"

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

prompt = "LLM에 대해 간단히 설명해주세요."
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=512
)
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]

Benchmark

model	ArenaHard	LogicKor	Kobest	Haerae	GSM8K-ko	Chinese text non-display rate
Qwen/Qwen2.5-3B-Instruct	23.28	6.0476	59.03	53.25	46.85	92.4
tryumanshow/Qwen2.5-3B-Instruct-ko	28.03	6.9643	63.72	51.79	50.57	94.9

• GSM8K from kuotient/gsm8k-ko
• Chinese text non-display rate
  • A metric to measure the degree of code-mixing.
  • Results measured from the top 1,000 entries of the CarrotAI/ko-instruction-dataset.

Key Highlights

This model was trained exclusively on private Korean datasets, resulting in a remarkable improvement in Korean instruction-following capabilities and human alignment performance. Interestingly, despite focusing solely on Korean data, the model also demonstrated significant enhancements in English benchmarks, such as ArenaHard. Overall, performance across various tasks has improved, though there was a slight drop in haerae task. Additionally, the training process partially resolved one of the longstanding issues in Qwen-2.5 series: the problem of code-mixing.

Limitation

One notable limitation is that the exclusive focus on natural language training has resulted in a decline in tool-calling performance. Addressing this issue will be a priority for future improvements to ensure a more well-rounded and versatile model.

Citation

@misc{qwen2.5,
    title = {Qwen2.5: A Party of Foundation Models},
    url = {https://qwenlm.github.io/blog/qwen2.5/},
    author = {Qwen Team},
    month = {September},
    year = {2024}
}

@article{qwen2,
      title={Qwen2 Technical Report}, 
      author={An Yang and Baosong Yang and Binyuan Hui and Bo Zheng and Bowen Yu and Chang Zhou and Chengpeng Li and Chengyuan Li and Dayiheng Liu and Fei Huang and Guanting Dong and Haoran Wei and Huan Lin and Jialong Tang and Jialin Wang and Jian Yang and Jianhong Tu and Jianwei Zhang and Jianxin Ma and Jin Xu and Jingren Zhou and Jinze Bai and Jinzheng He and Junyang Lin and Kai Dang and Keming Lu and Keqin Chen and Kexin Yang and Mei Li and Mingfeng Xue and Na Ni and Pei Zhang and Peng Wang and Ru Peng and Rui Men and Ruize Gao and Runji Lin and Shijie Wang and Shuai Bai and Sinan Tan and Tianhang Zhu and Tianhao Li and Tianyu Liu and Wenbin Ge and Xiaodong Deng and Xiaohuan Zhou and Xingzhang Ren and Xinyu Zhang and Xipin Wei and Xuancheng Ren and Yang Fan and Yang Yao and Yichang Zhang and Yu Wan and Yunfei Chu and Yuqiong Liu and Zeyu Cui and Zhenru Zhang and Zhihao Fan},
      journal={arXiv preprint arXiv:2407.10671},
      year={2024}
}