CLAUDE.md

You are a helpful AI scientist to build up the codebase for me.

This project is to train the open-sourced model to deploy CoT-like reasoning format on text-to-image and text-to-video generation quality assessment. You are using the LLaMAFactory to train the model, and write evaluation functions.

# Preparation

## Data

There are a folder called `ea-data/agent` and there are 3 subfolders:

* `vbench_results`: which stores the results for using proprietary models to evaluate different dimensions in vbench, and the results are CoT style.
* `t2i_results`: which stores the results for using proprietary models to evaluate different dimensions in T2I-CompBench, and the results are CoT style.
* `open_results`: which store the results for using proprietary models to evaluate open-ended queries.

Your first job is to write and execute the python script to clean the data in those aforementioned folders and convert them into the format align with `/home/data2/sltian/code/evaluation_agent_dev/LLaMA-Factory/data/alpaca_en_demo.json`.

如果指定， system 列对应的内容将被作为系统提示词。

history 列是由多个字符串二元组构成的列表，分别代表历史消息中每轮对话的指令和回答。注意在指令监督微调时，历史消息中的回答内容也会被用于模型学习。

指令监督微调数据集 格式要求 如下：

[
  {
    "instruction": "人类指令（必填）",
    "input": "人类输入（选填）",
    "output": "模型回答（必填）",
    "system": "系统提示词（选填）",
    "history": [
      ["第一轮指令（选填）", "第一轮回答（选填）"],
      ["第二轮指令（选填）", "第二轮回答（选填）"]
    ]
  }
]
下面提供一个 alpaca 格式 多轮 对话的例子，对于单轮对话只需省略 history 列即可。

[
  {
    "instruction": "今天的天气怎么样？",
    "input": "",
    "output": "今天的天气不错，是晴天。",
    "history": [
      [
        "今天会下雨吗？",
        "今天不会下雨，是个好天气。"
      ],
      [
        "今天适合出去玩吗？",
        "非常适合，空气质量很好。"
      ]
    ]
  }
]
对于上述格式的数据， dataset_info.json 中的 数据集描述 应为：

"数据集名称": {
  "file_name": "data.json",
  "columns": {
    "prompt": "instruction",
    "query": "input",
    "response": "output",
    "system": "system",
    "history": "history"
  }
}

## Train

After cleaning and collecting the data, you should write a script to train the `Qwen2.5-3B-Instruct` model using this created dataset.

The training is using `LLaMA-Factory`. You should read the dir and write a script to train the model.