qwen2.5-math系列模型生成问题

[Moyhub]

我这里使用qwen2.5-math-7b/72b-Instruct尝试生成数据，但是生成数据往往不可用。

1. 使用了qwen2.5的默认system。https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct
2. 也使用了自己编写的system prompt

截图如下：

[zhangchen-xu]

Sorry for the late reply as Github didn't give me issue notifications... 😔

I tried Qwen/Qwen2.5-Math-72B-Instruct and it works well from my side:

  {
    "id": 0,
    "pre_query_template": "<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n<|im_start|>user\n",
    "instruction": "Given the equation \\(x^2 + y^2 - 4x + 6y + 5 = 0\\), how can I determine the center and the radius of the circle it represents?",
    "response": null,
    "created": 1735442754,
    "gen_input_configs": {
      "temperature": 0.6,
      "top_p": 1.0,
      "input_generator": "Qwen/Qwen2.5-Math-72B-Instruct",
      "seed": null
    },
    "gen_response_configs": null
  },

You can add this to the model config:

  "Qwen/Qwen2.5-Math-72B-Instruct": {
    "model_name": "Qwen/Qwen2.5-Math-72B-Instruct",
    "stop_tokens": [
      "<|im_start|>",
      "<|im_end|>",
      "<|endoftext|>"
    ],
    "stop_token_ids": [
      151643,
      151644,
      151645
    ],
    "stop_tokens_assistant": [
      "Assistant",
      "assistant"
    ],
    "pre_query_template": "<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.<|im_end|>\n<|im_start|>user\n"
  },

I also checked tokenizer configs of Qwen2.5 math. It is weird... Maybe they use different system templates during the post training. Anyway, this system template for Qwen2 can work well in this case. You are Qwen, created by Alibaba Cloud. You are a helpful assistant.

I used Temperature=0.6. You may also need to decrease the temperature for this model family. I will release a dataset for Qwen2.5 Math once I have available GPUs.

[zhangchen-xu]

I will close this issue as we released Magpie-Align/Magpie-Qwen2.5-Math-Pro-300K-v0.1.