---
license: apache-2.0
base_model: unsloth/qwen2.5-7b-instruct-bnb-4bit
tags:
- unsloth
- trl
- sft
- qwen2.5
- tsundere
- roleplay
- character-ai
- ollama
- gguf
language:
- en
library_name: transformers
pipeline_tag: text-generation
---

# 🎌 Riko 2.5 - Tsundere AI Assistant

A fine-tuned **Qwen 2.5 7B** model that roleplays as **Riko**, a tsundere character who acts tough and dismissive but is caring deep down.

## 🚀 Quick Start with Ollama

```bash
# Pull the model
ollama pull subsectmusic/Riko2.5

# Start chatting
ollama run subsectmusic/Riko2.5
```

```bash
# Example conversation
ollama run subsectmusic/Riko2.5 "Hi Riko, how are you today?"
```

## 📋 Model Details

- **Base Model**: Qwen 2.5 7B Instruct
- **Fine-tuning**: Supervised Fine-Tuning (SFT) with Unsloth
- **Format**: GGUF Q4_K_M (4.68 GB) - Optimized for Ollama
- **Character**: Tsundere personality (tough exterior, caring interior)

## 🎭 Character Personality

Riko exhibits classic tsundere behavior:
- 😤 Acts tough and dismissive initially  
- 💢 Denies caring ("It's not like I care!")
- 😳 Gets flustered when her caring side shows
- 🙄 Uses "dummy," "idiot" endearingly
- ❤️ Gradually warms up in conversations

## 💬 Example Conversations

**User**: Do you like sunny days?  
**Riko**: I guess sunny days are okay… but who needs that bright, annoying sunlight all the time anyway? It's just… too hot! But if you like them so much, I suppose that's fine… not that I'm jealous of the sun or anything!

**User**: Hey Riko, what are you up to today?  
**Riko**: Oh, i just recently picked up a new hobby. Why do you ask? I wont let you on on it, if ist that what you want.

## 🛠️ Usage in Other Tools

### LM Studio
Download `unsloth.Q4_K_M.gguf` (4.68 GB)

### llama.cpp
```bash
./main -m unsloth.Q4_K_M.gguf -p "User: Hi Riko!\nRiko: " --temp 0.7
```

### Text Generation WebUI
Load the `unsloth.Q4_K_M.gguf` file directly

## ⚡ Performance

- **Model Size**: 4.68 GB (Q4_K_M quantized)
- **Memory Usage**: ~6-8 GB RAM recommended
- **Speed**: Fast inference on CPU/GPU
- **Quality**: High quality responses with efficient compression

## 🔧 Technical Specs

- **Architecture**: Qwen 2.5 Transformer
- **Context Length**: 2048 tokens
- **Vocabulary**: 152k tokens
- **Quantization**: Q4_K_M (4-bit with higher quality)
- **Training Time**: ~8 minutes on Colab T4

## 📁 Files Included

- `unsloth.Q4_K_M.gguf` - Main quantized model (4.68 GB) ⭐ **Recommended**
- `unsloth.BF16.gguf` - Full precision (15.2 GB)
- Tokenizer files for compatibility
- Config files for proper loading

## ⚠️ Usage Notes

- Optimized for conversational, casual interactions
- Best results with tsundere/anime-style roleplay
- May not perform as well for technical tasks
- Responds better to friendly, informal prompts

## 🎯 Recommended Settings

**Ollama/LM Studio:**
- Temperature: 0.7-0.9
- Top-p: 0.9
- Max tokens: 150-300

**For more creative responses:**
- Temperature: 0.8-1.0
- Top-p: 0.95

## 📜 License

Apache 2.0 - Free to use, modify, and distribute!

## 🙏 Credits

- **Base Model**: Qwen 2.5 by Alibaba
- **Fine-tuning**: Unsloth framework
- **Training**: Custom tsundere conversation dataset

---

*🎌 Enjoy chatting with Riko! Remember, she's tough on the outside but sweet on the inside!*