README.md

---
license: apache-2.0
datasets:
- BelleGroup/train_0.5M_CN
language:
- en
- zh
tags:
- text-generation-inference
widget:
- text: |-
    <|im_start|>user
    请以『春天的北京』为题写一首诗歌
    <|im_end|>
    <|im_start|>assistant
  example_title: generation zh
---
# Baichuan 7B ChatML

## 介绍 Introduction

`baichuan-7B-chatml` 是支持多轮对话兼容于 ChatML 的模型。
模型基于 [baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B) 微调而成。
`baichuan-7B-chatml` 模型支持商用。但按照baichuan-7B的要求，如果将baichuan-7B衍生品用作商业用途，需要联系[baichuan-7B 的许可方](https://github.com/baichuan-inc/baichuan-7B#%E5%8D%8F%E8%AE%AE)。

__需要注意：在面对事实性知识任务时，模型可能会生成不正确的信息或者产生不稳定的输出（有时可以返回正确答案，有时不能）。__

`baichuan-7B-chatml` is a model that supports multi-turn dialog and is compatible with ChatML.
The model is fine-tuned based on [baichuan-7B](https://huggingface.co/baichuan-inc/baichuan-7B).
`baichuan-7B-chatml` model supports commercial use. However, according to the requirements of baichuan-7B, if baichuan-7B derivatives are used for commercial purposes, you need to contact [baichuan-7B](https://github.com/baichuan-inc/baichuan-7B#%E5%8D%8F%E8%AE%AE)。

__Note: When dealing with factual knowledge tasks, it may generate incorrect information or unstable output (sometimes it can return the correct answer, sometimes not).__

## 代码示例 Examples

模型在百川的基础上提供了对轮对话的函数供调用。

The model provides a function for multi-turn dialogs.

```ipython
>>> from transformers import AutoTokenizer, AutoModelForCausalLM
>>> tokenizer = AutoTokenizer.from_pretrained("tibok/baichuan-7B-chatml", trust_remote_code=True)
>>> model = AutoModelForCausalLM.from_pretrained("tibok/baichuan-7B-chatml", device_map="auto", trust_remote_code=True)
>>> response, history = model.chat(tokenizer, "请以『春天的北京』为题写一首诗歌", history=[])
春天的北京，
花开万丈，
春意盎然，
清风送暖。
<|im_end|>
>>> response, history = model.chat(tokenizer, "能不能再写一首关于香山的？", history=history)
>>> print(response)
香山之巅，
芳草连天。
清泉潺潺，
山峦绵绵。
<|im_end|>
```

## 更多细节 Details

- Dataset： [BelleGroup/train_0.5M_CN](https://huggingface.co/datasets/BelleGroup/train_0.5M_CN) 
- steps: 13800
- batch_size: 8
- seq_len： 2048