---
license: apache-2.0
base_model: FlagAlpha/Llama3-Chinese-8B-Instruct
datasets:
  - Moemu/Muice-Dataset
tags:
- chinese
- lora
- llama3
library_name: peft
language:
  - zh
---

## 模型说明

基于 [FlagAlpha/Llama3-Chinese-8B-Instruct](https://huggingface.co/FlagAlpha/Llama3-Chinese-8B-Instruct) 微调而成的沐雪角色扮演模型，在主动对话上表现出一定的优势（固定Prompt）。使用本模型需要模型本体。

沐雪设定可参见：[Moemu/Muice-Chatbot](https://github.com/Moemu/Muice-Chatbot?tab=readme-ov-file#沐雪人设)

部分训练集参见：[Moemu/Muice-Dataset](https://huggingface.co/datasets/Moemu/Muice-Dataset)

## 对话示例

![此对话示例来自于沐雪直播间](https://img.snowy.moe/2024/image-20250126164517960.png)

## System Prompt

在不同任务中，模型拥有不同的System Prompt。具体参见：[Muice-Chatbot/llm/utils/auto_system_prompt.py at main · Moemu/Muice-Chatbot](https://github.com/Moemu/Muice-Chatbot/blob/main/llm/utils/auto_system_prompt.py)

## 评估

| 模型名                                          | 新话题发起分数 | 直播对话性能 | 日常聊天性能 | 综合对话分数 |
| ----------------------------------------------- | -------------- | ------------ | ------------ | ------------ |
| Muice-2.3-chatglm2-6b-int4-pt-128-1e-2          | 2.80           | 4.00         | 4.33         | 3.45         |
| Muice-2.4-chatglm2-6b-int4-pt-128-1e-2          | 3.20           | 4.00         | 3.50         | 3.45         |
| Muice-2.4-Qwen2-1.5B-Instruct-GPTQ-Int4-2e-3    | 1.40           | 3.00         | 6.00         | 5.75         |
| Muice-2.5.3-Qwen2-1.5B-Instruct-GPTQ-Int4-2e-3  | 4.04           | 5.00         | 4.33         | 5.29         |
| Muice-2.6.2-Qwen-7B-Chat-Int4-5e-4              | 5.20           | 5.67         | 4.00         | 5.75         |
| Muice-2.7.0-Qwen-7B-Chat-Int4-1e-4              | 2.40           | 5.30         | 6.00         | \            |
| Muice-2.7.1-Qwen2.5-7B-Instruct-GPTQ-Int4-8e-4  | **5.40**       | 4.60         | 4.50         | **6.76**     |
| **Muice-2.7.1-Llama3_Chinese_8b_Instruct-8e-5** | 4.16           | **6.34**     | **8.16**     | 6.20         |

## 训练参数

- learning_rate: 8e-05
- train_batch_size: 4
- eval_batch_size: 4
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- num_epochs: 5.0
- mixed_precision_training: Native AMP