历史消息貌似没有正确加到训练数据中

[ylsdamxssjxxdd]

Reminder

• I have read the README and searched the existing issues.

System Info

root@8d9356cd571a:/nerv/nerv-workspace# llamafactory-cli env
[2024-07-04 16:14:35,683] [INFO] [real_accelerator.py:203:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[WARNING] async_io requires the dev libaio .so object and headers but these were not found.
[WARNING] async_io: please install the libaio-dev package with apt
[WARNING] If libaio is already installed (perhaps from source), try setting the CFLAGS and LDFLAGS environment variables to where it can be found.
[WARNING] Please specify the CUTLASS repo directory as environment variable $CUTLASS_PATH
[WARNING] sparse_attn requires a torch version >= 1.5 and < 2.0 but detected 2.3
[WARNING] using untested triton version (2.3.1), only 1.0.0 is known to be compatible

• llamafactory version: 0.8.3.dev0
• Platform: Linux-5.10.0-8-generic-x86_64-with-glibc2.35
• Python version: 3.10.12
• PyTorch version: 2.3.1+cu121 (GPU)
• Transformers version: 4.42.3
• Datasets version: 2.20.0
• Accelerate version: 0.32.1
• PEFT version: 0.11.1
• TRL version: 0.9.4
• GPU type: NVIDIA GeForce RTX 3090 Ti
• DeepSpeed version: 0.14.4
• vLLM version: 0.5.0.post1

Reproduction

• 我修改了LLaMA-Factory-main/src/llamafactory/data/loader.py中的print_function(next(iter(dataset)))为

for item in dataset:
    print_function(item)

这样应该可以打印出所有训练数据

• 我的数据集就是，一个字没改~

[
  {
    "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）"]
    ]
  }
]

• 训练qwen2时weiui输出

input_ids:
[151644, 8948, 198, 72448, 45139, 99689, 9909, 30767, 68756, 7552, 151645, 198, 151644, 872, 198, 99363, 99620, 109504, 9909, 30767, 68756, 7552, 151645, 198, 151644, 77091, 198, 99363, 99620, 102104, 9909, 30767, 68756, 7552, 151645]
inputs:
<|im_start|>system
系统提示词（选填）<|im_end|>
<|im_start|>user
第一轮指令（选填）<|im_end|>
<|im_start|>assistant
第一轮回答（选填）<|im_end|>
label_ids:
[-100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, 99363, 99620, 102104, 9909, 30767, 68756, 7552, 151645]
labels:
第一轮回答（选填）<|im_end|>

Expected behavior

按道理weiui应该输出
第一轮回答
第二轮回答
模型回答
但是现在只有第一轮回答，说明历史消息没有正确加到训练数据中~

Others

No response

[hiyouga]

抱歉 已修复

[maksimstw]

没太理解，这意味着在这个bug修复前，模型训练的时候都只是训练在第一轮上？之后几轮的数据完全没用上？请问这个bug存在多久了，是指到两周前才修好吗？

[hiyouga]

@maksimstw 在被修复为止存在了7天时间