Edit model card

Updates

  • [June 04, 2024] 🌈 Add more Chinese training data to improve Chinese performance
  • [May 23, 2024] 🔥 Support function calling

Model Summary

Gemma-1.1-7B-Chinese-Chat is an instruction-tuned language model for Chinese & English users built upon the gemma-1.1-7b-it model.(Github)

Developed by: ycjcl868

  • License: Gemma License
  • Base Model: gemma-1.1-7b-it
  • Model Size: 8.54B
  • Context length: 8K

Introduction

This is the first model specifically fine-tuned for Chinese & English user through SFT based on the gemma-1.1-7b-it model.

Compared to the original gemma-1.1-7b-it model, our Gemma-1.1-7B-Chinese-Chat model significantly reduces the issues of "Chinese questions with English answers" and the mixing of Chinese and English in responses.

Training details:

  • epochs: 3
  • learning rate: 0.0001
  • learning rate scheduler type: cosine
  • Warmup ratio: 0.1
  • cutoff len (i.e. context length): 8192
  • global batch size: 2

Usage

Chat Template

The instruction-tuned models use a chat template that must be adhered to for conversational use. The easiest way to apply it is using the tokenizer's built-in chat template, as shown in the following snippet.

Let's load the model and apply the chat template to a conversation. In this example, we'll start with a single user interaction:

from transformers import AutoTokenizer, AutoModelForCausalLM
import transformers
import torch

model_id = "ycjcl868/Gemma-1.1-7B-Chinese-Chat"

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

messages = [
    { "role": "user", "content": "写一首诗" },
]
input_ids = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=8192,
    do_sample=True,
    temperature=0.6,
    top_p=0.9,
)
response = outputs[0][input_ids.shape[-1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

# 春风拂柳,春雨轻轻,
# 花香满园,鸟鸣清脆。
# 一朵红梅,红彤彤的,
# 美不胜收,美不胜收。

Examples

User: 参加四年级数学竞赛同学中,男同学获奖人数比女同学多2人,女同学获奖人数比男同学人数的一半多2人,男女同学各有多少人获奖?
Assistant:
设男同学的获奖人数
为x,则女同学的获奖人数为y,所以有以下两式:
x=y+2
y=2(x+2)
将第一式代入第二式中,得:
y=2(y+2)+2
y=2y+4+2
y=2y+6
2y=y+6
y=6
因此,女同学获奖人数为6人,男同学获奖人数为8人。
User: 我的蓝牙耳机坏了,我应该去看耳鼻喉科还是牙科?
Assistant: 蓝牙耳机是属于电子产品的,而不是耳鼻喉科或牙科的范畴。如果你怀疑耳机有故障,建议去电子产品维修店或者原厂维修中心进行检查。如果需要新的耳机,你也可以去电器商店或电子产品市场购买。
User: 来几句成语
Assistant: “一日三尺”是指过时间一切都会改变。
User: 列出5种可以改善睡眠质量的方法
Assistant:
1. 保持规律的睡眠时间表
2. 避免下午摄入咖啡因
3. 避免睡前摄入酒精
4. 避免睡前摄入高糖饮食
5. 每天至少进行30分钟的运动
User: 回答一个开放性的问题:人类存在的意义是什么?
Assistant: 人类存在的意义是一个复杂的哲学和哲学问题,因为人类存在于自然界中,我们作为人类有责任和义务来保护地球和生物,而我们作为人类也有责任来发展自己的精神和思想,来创造更好的未来。
我从哲学角度来看,人类存在的意义是通过我们自己的思想和行动,来实现我们对精神和精神的追求。通过我们自己的努力和创造力,我们可以创造出更多的艺术、文化和思想,来促进人类的进步和发展。
User: 解释以下缩写: JSON、CEO、PDF、HTML、NASA
Assistant:
JSON(JavaScript Object Notation):JavaScript对象的表示法,用来在JavaScript和许多其他编程语言之间传递数据。
CEO(Chief Executive Officer):首席执行官,指企业的高级管理人员,通常是指企业的主席或总裁。
PDF(Portable Document Format):可移植文档格式,是一种电子文档格式,可以确保文档在不同的操作系统���设备上都能以一致的方式呈现。
HTML(HyperText Markup Language):超文本标记语言,是网页内容的标记语言,用来定义网页的结构和内容。
NASA(National Aeronautics and Space Administration):美国国家航空航天局,是美国政府的宇航机构,负责美国在太空和航空方面的研究和发展。

Function call

User:

以下是您可以使用的工具列表:

```python
def internet_search(query: str):
    \"\"\"
    Returns a list of relevant document snippets for a textual query retrieved from the internet

    Args:
        query (str): Query to search the internet with
    \"\"\"
    pass
```

```python
def directly_answer():
    \"\"\"
    Calls a standard (un-augmented) AI chatbot to generate a response given the conversation history
    \"\"\"
    pass
```

写 'Action:',后跟要调用的 JSON 中的操作列表,例如.
Action:
```json
[
    {
        "tool_name": "tool name (one of [internet_search, directly_answer])",
        "parameters": "the input to the tool"
    }
]
```

帮我找到今天的新闻有哪些:

Response:

Action:
[
  {
    "tool_name": "internet_search", 
    "parameters": "今天有哪些新闻"
  }
]
Downloads last month
7
Safetensors
Model size
8.54B params
Tensor type
BF16
·
Inference Examples
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social visibility and check back later, or deploy to Inference Endpoints (dedicated) instead.