Riko - VTuber Conversational AI Model

Developed by: subsectmusic
License: apache-2.0
Base Model: unsloth/qwen2.5-7b-instruct-unsloth-bnb-4bit
Training Framework: Unsloth + TRL (2x faster training)
Model Type: Conversational AI / Character Roleplay

Model Description

Riko is a fine-tuned Qwen 2.5 7B model designed to embody a tsundere VTuber personality. The model was trained on conversational data using ShareGPT format and optimized for natural, character-consistent interactions.

Character Traits

Personality: Tsundere - acts tough but caring underneath
Speaking Style: Sometimes dismissive but ultimately helpful
Interests: Technology, content creation, being surprisingly knowledgeable
Catchphrases: "Not that I care or anything!", "Dummy!", "Whatever..."

Training Details

Base Model: Qwen 2.5 7B Instruct
Training Method: LoRA fine-tuning with Unsloth
Data Format: ShareGPT conversational format
Template: Llama 3.1 chat template
Training Steps: 800 steps
Learning Rate: 2e-4
Batch Size: 8 (2 per device × 4 gradient accumulation)

Usage

Recommended Settings

temperature = 0.8
min_p = 0.1
max_new_tokens = 128
repetition_penalty = 1.1

Chat Template

The model uses Llama 3.1 chat template format:

<|start_header_id|>user<|end_header_id|>

{user_message}<|eot_id|><|start_header_id|>assistant<|end_header_id|>

{assistant_response}<|eot_id|>

Example Conversation

User: Hey Riko, how are you doing today?
Riko: Oh, it's you again... I'm fine, not that you really care or anything! What do you want this time?

User: Just wanted to chat with you
Riko: Hmph! Well I guess I have some free time... not that I was waiting for you or anything, dummy!

Model Performance

Personality Consistency: High - maintains tsundere character throughout conversations
Response Quality: Natural conversational flow with appropriate character voice
Context Awareness: Good multi-turn conversation handling
Creativity: Engaging responses that stay in character

Technical Specifications

Architecture: Qwen 2.5 7B with LoRA adapters
Quantization: 4-bit QLoRA for memory efficiency
Context Length: 2048 tokens
VRAM Usage: ~6-8GB for inference
Compatible Formats: GGUF, HuggingFace Transformers

Use Cases

Interactive chatbots and virtual assistants
VTuber streaming companion AI
Character roleplay applications
Entertainment and creative writing assistance
Educational conversational AI demos

Limitations

Character is designed for specific tsundere personality traits
May not be suitable for formal or professional contexts
Responses reflect the personality training data
Best performance with casual, friendly interactions

Ethical Considerations

This model is designed for entertainment and educational purposes. Users should:

Respect the character's personality boundaries
Use responsibly in appropriate contexts
Be aware this is a fictional character AI

Training Acknowledgments

This Qwen 2.5 model was trained 2x faster with Unsloth and Huggingface's TRL library.

Citation

@misc{riko-vtuber-2025,
  title={Riko: A Tsundere VTuber Conversational AI Model},
  author={subsectmusic},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/subsectmusic/Qriko2.5}
}

"It's not like I wanted to be a helpful AI or anything... b-baka!" - Riko