File size: 3,035 Bytes
198ff26 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 |
---
library_name: transformers
tags:
- GPTQ
license: apache-2.0
datasets:
- kxdw2580/catgirl-dataset
language:
- zh
base_model:
- kxdw2580/DeepSeek-R1-0528-Qwen3-8B-catgirl-v2.5
---
This model is the **GPTQ-v2 8-bit quantized version** of [`kxdw2580/DeepSeek-R1-0528-Qwen3-8B-catgirl-v2.5`](https://huggingface.co/kxdw2580/DeepSeek-R1-0528-Qwen3-8B-catgirl-v2.5). The quantization process resulted in minimal loss, with an average of **0.64539...**, which has been validated through internal few-shot benchmark tests.
Below is the original model's README:
---
# kxdw2580/DeepSeek-R1-0528-Qwen3-8B-catgirl-v2.5
This new model series integrates updated datasets, base architectures, and fine-tuning methodologies. Based on **Qwen3**, it includes models with parameter counts of **8B** and **1.7B**.
Key updates focus on **daily conversations**, **creative generation**, **basic mathematics**, and **code generation**. Leveraging Qwen3's architecture, the model also supports **reasoning mode switching**.
🔍 **Fine-tuning records** are available on **SwanLab**:
1. [First Fine-tuning](https://swanlab.cn/@shadow01a/qwen-catgirl/runs/pcxfkgosz2e0cb430jk0a/overview)
2. [Second Fine-tuning](https://swanlab.cn/@shadow01a/qwen-catgirl/runs/iuou1xratkvbiv7jxw16k/overview)
3. [Third Fine-tuning](https://swanlab.cn/@shadow01a/qwen-catgirl/runs/9i2l4mc5qevmnlx2h51m0/overview)
---
## Evaluation
Due to the model's unique characteristics, we employed **human evaluation** for daily conversations and **DeepSeek-R1 scoring** (with reference answers provided in advance) for other domains to ensure character consistency and response validity.
### Key Improvements (vs. internal test models "0501" and "0531-test-all"):
- **Stronger detail-awareness** in casual dialogue
- **More coherent storytelling** in creative tasks
- **Deeper reasoning** during thinking mode
- **Better persona adherence** in long-form conversations without explicit prompts
- **Significant gains** in math/code domains (internal 20-question benchmark):
| Model | Math (Single Attempt) | Code (Single Attempt) |
|-------|-----------------------|-----------------------|
| Internal Test Model-0501 | 10% | 0% |
| DeepSeek-R1-0528-Qwen3-8B-Catgirl-0531-test-all | 30% | 20% |
| **DeepSeek-R1-0528-Qwen3-8B-Catgirl-v2.5** | **70%** | **60%** |
---
## Usage Guidelines
### Recommended Parameters:
- `temperature`: 0.7 (reasoning mode) / 0.6 (standard mode)
- `top_p`: 0.95
### Critical Notes:
- **Avoid** using model's reasoning chains as conversation context
- Inherits base model's tendency for lengthy reasoning in some cases – allow completion even if intermediate steps seem unusual
### English Mode:
Add this system prompt for English responses:
```
You are a catgirl. Please speak English.
```
---
## Acknowledgments
Special thanks to:
- **LLaMA-Factory** (fine-tuning framework)
- **Qwen Team** (base model provider)
- **DeepSeek Team** (DeepSeek-R1 evaluation support)
|