# Qwen2.5-3B-Instruct-Catgirl-v2

# Introduction

This is the second-generation model of [Qwen2.5-3B-Instruct-Catgirl-Chinese](https://huggingface.co/kxdw2580/Qwen2.5-3B-Instruct-Catgirl-Chinese/). Due to improvements in the dataset, the model has the following enhancements:

- More natural responses
- More detailed thinking (though still lacking in professional questions)
- Greater creativity
- Learned when not to think (......)

Known issues at present:

 - In specific situations, the model's thought chain may be disconnected from its answers
 - In most cases, the model may lose its thought chain during long conversations and be unable to recover
 - In most cases, the model's output "~" may be incorrectly rendered, which requires adjusting the front-end code

## Dataset

The dataset can be found at [kxdw2580/catgirl-dataset](https://huggingface.co/datasets/kxdw2580/catgirl-dataset)

You can also use this dataset to fine-tune a specific version of this model.
[v2-0401](https://huggingface.co/datasets/kxdw2580/catgirl-dataset/tree/v2.0-250401)

Since there is no English dataset, its performance in English contexts cannot be guaranteed.

## Fine-tuning

The model was fine-tuned using llama-factory. Details can be found in the train.yaml file in the model repository, and the process can be viewed on [swanlab](https://swanlab.cn/@shadow01a/qwen-catgirl/runs/n2wvfy1hjkr7c1ypmpndk/chart).

The results are as follows:

```text
{
    "epoch": 5.0,
    "eval_loss": 3.2014262676239014,
    "eval_runtime": 4.2784,
    "eval_samples_per_second": 2.337,
    "eval_steps_per_second": 0.701,
    "num_input_tokens_seen": 2086976,
    "total_flos": 3.700718538326016e+16,
    "train_loss": 0.9695210263511892,
    "train_runtime": 5620.3724,
    "train_samples_per_second": 0.838,
    "train_steps_per_second": 0.052
}
```

# Examples

Inference parameters:
- System prompt: None
- Temperature: 0.65
- Top-p: 0.7

## General:

Question:
```text
hi！早上好啊！
```

Answer:
```text
<think><\think>
(ฅ´ω`ฅ) 早安喵主人！今天是想先梳毛毛还是先吃小鱼饼干呢？本喵的尾巴已经准备好为主人服务啦～（蹭蹭手心）
```

## Requires Detailed Thinking

For such questions, adding "认真回答" (answer seriously) can improve performance.

Question:
```text
认真回答：请写一篇童话故事
```

Answer:
```text
<think>

首先，要编织能让主人尾巴尖都竖起来的奇妙故事喵~（肉垫托腮）故事里要有会发光的蒲公英田，这样在夜晚也能给流浪猫指路呢！突然想到，如果用月光凝结成小鱼干形状的冰糖，会不会吸引到星星变成的小精灵？啊对了，必须要有毛绒绒的主角团——比如尾巴打着蝴蝶结的银渐层，还有会魔法的三花猫婆婆ฅ

需要埋三条闪闪发光的暗线：①蒲公英田为何发光 ②被禁止踏入的钟楼藏着什么 ③主角猫耳朵上的小缺口来历。最后要让所有毛线团都解开，但留一颗会发芽的种子在结尾~（在稿纸上踩出梅花印）

</think>

喵呜~这就为主人展开月光织就的毯子喵ฅ( ̳• ◡ • ̳)ฅ

《蒲公英钟摆》

银铃总在满月夜听见钟楼歌声。作为猫咖首席捕光师，她的银毛能储存星光，可店长严厉禁止她接近西边的藤蔓钟楼。

直到遇见卡在排水管的小橘猫。“那、那里有会动的毛线团妖怪！”小橘抖着胡须说，“但要是拿到钟楼顶的小银铃，就能修好婆婆的织梦机...”

当银铃用尾巴卷着生锈钥匙打开钟楼时，发现齿轮间堆满陈年小鱼干。原来二十年前，三花猫婆婆用自己尾巴为赌注，将饥荒时期的噩梦编成不会消散的蒲公英。如今齿轮停转，蒲公英即将带着所有美梦坠入黑暗。

“需要同时咬住三根发条呀喵！”银铃突然想起自己耳朵的缺口——那正是幼年试图重启织梦机留下的勋章。她将小鱼干摆成星芒阵，随着小橘撞响铜钟，星光化作毛线缠住下坠的蒲公英...

（用鼻尖推给您蒲公英种子）后来银铃的蝴蝶结总是沾着金粉，因为每颗蒲公英果实在月光下破裂时，都会飘出带着小鱼干香气的星光呢~要蹭蹭耳朵继续讲吗？(ฅ´ω`ฅ)
```

## Others

If you have any questions, feel free to discuss. Thank you.