README.md

---
license: apache-2.0
language:
- en
- zh
base_model: meta-llama/Meta-Llama-3-8B-Instruct
tags: 
- text-generation
- transformers
- lora
- llama.cpp
- autoawq
- auto-gptq
datasets:
- llamafactory/alpaca_zh
- llamafactory/alpaca_gpt4_zh
---

# Meta-Llama-3-8B-Instruct-zh-10k: A Llama🦙 which speaks Chinese / 一只说中文的羊驼🦙

## Model Details / 模型细节
This model, <u>`Meta-Llama-3-8B-Instruct-zh-10k`</u>, was fine-tuned from the original [Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) due to its underperformance in Chinese. Utilizing the LoRa technology within the [LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory) utilities, this model was adapted to better handle Chinese through three epochs on three corpora: `alpaca_zh`, `alpaca_gpt4_zh`, and `oaast_sft_zh`, amounting to approximately 10,000 examples. This is reflected in the `10k` in its name.

由于原模型[Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct)在中文上表现欠佳，于是该模型 <u>`Meta-Llama-3-8B-Instruct-zh-10k`</u> 微调自此。在[LLaMA-Factory](https://github.com/hiyouga/LLaMA-Factory)工具下，利用LoRa 技术，通过`alpaca_zh`、`alpaca_gpt4_zh`和`oaast_sft_zh`三个语料库上、经过三个训练轮次，我们将该模型调整得更好地掌握了中文。三个语料库共计约10,000个样本，这也是其名字中的 `10k` 的由来。

For efficient inference, the model was converted to the gguf format using [llama.cpp](https://github.com/ggerganov/llama.cpp) and underwent quantization, resulting in a compact model size of about 3.18 GB, suitable for distribution across various devices.

为了高效的推理，使用 [llama.cpp](https://github.com/ggerganov/llama.cpp)，我们将该模型转化为了gguf格式并量化，从而得到了一个压缩到约 3.18 GB 大小的模型，适合分发在各类设备上。

### LoRa Hardware / LoRa 硬件
- RTX 4090D x 1

> [!NOTE]
> The complete fine-tuning process took approximately 12 hours. / 完整微调过程花费约12小时。

Additional fine-tuning configurations are avaiable at [Hands-On LoRa](https://github.com/XavierSpycy/hands-on-lora) or [Llama3Ops](https://github.com/XavierSpycy/llama-ops).

更多微调配置可以在我的个人仓库 [Hands-On LoRa](https://github.com/XavierSpycy/hands-on-lora) 或 [Llama3Ops](https://github.com/XavierSpycy/llama-ops) 获得。

### Other Models / 其他模型
- <u>llama.cpp</u>
    - [Meta-Llama-3-8B-Instruct-zh-10k-GGUF](https://huggingface.co/XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k-GGUF)

- <u>AutoAWQ</u>
    - [Meta-Llama-3-8B-Instruct-zh-10k-AWQ](https://huggingface.co/XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k-AWQ)

- <u>AutoGPTQ</u>
    - [Meta-Llama-3-8B-Instruct-zh-10k-GPTQ](https://huggingface.co/XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k-GPTQ)

### Model Developer / 模型开发者
- **Pretraining**: Meta
- **Fine-tuning**: [XavierSpycy @ GitHub ](https://github.com/XavierSpycy) | [XavierSpycy @ 🤗](https://huggingface.co/XavierSpycy)

- **预训练**: Meta
- **微调**: [XavierSpycy @ GitHub](https://github.com/XavierSpycy) | [XavierSpycy @ 🤗 ](https://huggingface.co/XavierSpycy)


### Usage / 用法
This model can be utilized like the original <u>Meta-Llama3</u> but offers enhanced performance in Chinese.

我们能够像原版的<u>Meta-Llama3</u>一样使用该模型，而它提供了提升后的中文能力。

#### 1. How to use in transformers
```python
# !pip install accelerate

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k"

model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

prompt = "你好，你是谁？"

messages = [
    {"role": "system", "content": "你是一个乐于助人的助手。"},
    {"role": "user", "content": prompt}]

input_ids = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

terminators = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|eot_id|>")]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    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))
# 我是一个人工智能助手，旨在帮助用户解决问题和完成任务。
# 我是一个虚拟的人工智能助手，能够通过自然语言处理技术理解用户的需求并为用户提供帮助。
```

#### 2. How to use in llama.cpp / 如何在llama.cpp中使用


```python
# CMAKE_ARGS="-DLLAMA_BLAS=ON -DLLAMA_BLAS_VENDOR=OpenBLAS # -DLLAMA_CUDA=on" \
#  pip install llama-cpp-python \
#  --extra-index-url https://abetlen.github.io/llama-cpp-python/whl/cu121

# Please download the model weights first. / 请先下载模型权重。

from llama_cpp import Llama

llm = Llama(
      model_path="/path/to/your/model/Meta-Llama-3-8B-Instruct-zh-10k-GGUF/meta-llama-3-8b-instruct-zh-10k.Q8_0.gguf",
      n_gpu_layers=-1)

# Alternatively / 或者
# llm = Llama.from_pretrained(
#     repo_id="XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k-GGUF",
#     filename="*Q8_0.gguf",
#     verbose=False
# )

output = llm(
      "Q: 你好，你是谁？A:", # Prompt
      max_tokens=256, # Generate up to 32 tokens, set to None to generate up to the end of the context window
      stop=["Q:", "\n"], # Stop generating just before the model would generate a new question
      echo=True # Echo the prompt back in the output
) # Generate a completion, can also call create_completion

print(output['choices'][0]['text'].split("A:")[1].strip())

# 我是一个人工智能聊天机器人，我的名字叫做“智慧助手”，我由一群程序员设计和开发的。我的主要任务就是通过与您交流来帮助您解决问题，为您提供相关的建议和支持。
```

#### 3. How to use with AutoAWQ / 如何与AutoAWQ一起使用
```python
# !pip install autoawq

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k-AWQ"

model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

prompt = "你好，你是谁？"

messages = [
    {"role": "system", "content": "你是一个乐于助人的助手。"},
    {"role": "user", "content": prompt}]

input_ids = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

terminators = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|eot_id|>")]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    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))
# 你好！我是一个人工智能助手，我的目的是帮助人们解决问题，回答问题，提供信息和建议。
```

#### 4. How to use with AutoGPTQ / 如何与AutoGPTQ一起使用
```python
# !pip install auto-gptq --no-build-isolation

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "XavierSpycy/Meta-Llama-3-8B-Instruct-zh-10k-GPTQ"

model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_id)

prompt = "什么是机器学习？"

messages = [
    {"role": "system", "content": "你是一个乐于助人的助手。"},
    {"role": "user", "content": prompt}]

input_ids = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

terminators = [tokenizer.eos_token_id, tokenizer.convert_tokens_to_ids("<|eot_id|>")]

outputs = model.generate(
    input_ids,
    max_new_tokens=256,
    eos_token_id=terminators,
    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))
# 机器学习是人工智能（AI）的一个分支，它允许计算机从数据中学习并改善其性能。它是一种基于算法的方法，用于从数据中识别模式并进行预测。机器学习算法可以从数据中学习，例如文本、图像和音频，并从中获得知识和见解。
```

Further details about the deployment are available in the GitHub repository [Llama3Ops: From LoRa to Deployment with Llama3](https://github.com/XavierSpycy/llama-ops).

更多关于部署的细节可以在我的个人仓库 [Llama3Ops: From LoRa to Deployment with Llama3](https://github.com/XavierSpycy/llama-ops) 获得。

## Ethical Considerations, Safety & Risks / 伦理考量、安全性和风险
Please refer to [Meta Llama 3's Ethical Considerations](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct#ethical-considerations-and-limitations) for more information. Key points include bias monitoring, responsible usage guidelines, and transparency in model limitations.

请参考 [Meta Llama 3's Ethical Considerations](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct#ethical-considerations-and-limitations)，以获取更多细节。关键点包括偏见监控、负责任的使用指南和模型限制的透明度。

## Limitations / 局限性
- The comprehensive abilities of the model have not been fully tested. 

- While it performs smoothly in Chinese conversations, further benchmarks are required to evaluate its full capabilities. The quality and quantity of the Chinese corpora used may also limit model outputs. 

- Based on current observations, it fundamentally meets the standards in common sense, logic, sentiment analysis, safety, writing, code, and function calls. However, there is room for improvement in role-playing, mathematics, and handling complex tasks with the same text but different meanings.

- Additionally, catastrophic forgetting in the fine-tuned model has not been evaluated.

- 该模型的全面的能力尚未全部测试。

- 尽管它在中文对话中表现流畅，但需要更多的测评以评估其完整的能力。中文语料库的质量和数量可能都会对模型输出有所制约。

- 根据目前的观察，它在常识、逻辑、情绪分析、安全性、写作、代码和函数调用上基本达标，然而，在角色扮演、数学、复杂的同文异义等任务上有待提高。

- 另外，微调模型中的灾难性遗忘尚未评估。

## Acknowledgements / 致谢
We thank Meta for their open-source contributions, which have greatly benefited the developer community, and acknowledge the collaborative efforts of developers in enhancing this community. 

我们感谢 Meta 的开源贡献，这极大地帮助了开发者社区，同时，也感谢致力于提升社区的开发者们的努力。

## References / 参考资料

```
@article{llama3modelcard,
title={Llama 3 Model Card},
author={AI@Meta},
year={2024},
url = {https://github.com/meta-llama/llama3/blob/main/MODEL_CARD.md}}

@inproceedings{zheng2024llamafactory,
  title={LlamaFactory: Unified Efficient Fine-Tuning of 100+ Language Models},
  author={Yaowei Zheng and Richong Zhang and Junhao Zhang and Yanhan Ye and Zheyan Luo and Zhangchi Feng and Yongqiang Ma},
  booktitle={Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations)},
  address={Bangkok, Thailand},
  publisher={Association for Computational Linguistics},
  year={2024},
  url={http://arxiv.org/abs/2403.13372}}
```