🦊 QwRiko3-4B-Instruct-2507 — Tsundere Kitsune AI (GGUF • Ollama • Tools)

📋 Model Overview

QwRiko3-4B-Instruct-2507 is a conversational AI model fine-tuned to embody Riko, a tsundere kitsune character. This release targets GGUF for Ollama first, with solid tool calling support when run via Ollama’s tools API. A PyTorch build (Transformers) is also supported.

• Model ID (this repo): subsectmusic/qwriko3-4b-instruct-2507
• Primary format: GGUF (Ollama-compatible)
• Alt format: PyTorch (Transformers)
• Base Model: Qwen/Qwen3-4B-Instruct
• Parameters: ~4B
• License: Apache-2.0 (repo)
• Project: Project Horizon LLM
• Developer: @subsectmusic
• Training Framework: Unsloth + TRL (SFT)

🎭 Character Profile: Riko

• Tsundere cadence: “It’s not like I like you or anything… b-baka!”
• Kitsune vibes: fox-spirit mischief + sly wisdom
• Emotional core: tough shell, soft center
• Style: snappy, teasing, ultimately caring

---

🚀 Quick Start (Ollama • GGUF)

These steps assume you have a local GGUF file named qwriko3-4b-instruct-2507.Q4_K_M.gguf in the working directory. If your filename differs, update the FROM path in the Modelfile accordingly.

1. Create a Modelfile (exact content below is also saved as Modelfile in this package):

# Modelfile
FROM ./qwriko3-4b-instruct-2507.Q4_K_M.gguf
PARAMETER num_ctx 8192
# (Optional) you can set temperature/top_p/etc. via `ollama run -p` or the API.

2. Create the Ollama model:

ollama create qwriko3-4b-instruct-2507 -f Modelfile

3. Chat:

ollama run qwriko3-4b-instruct-2507 "Riko, give me a playful hello."

Tool Calling with Ollama (cURL)

curl http://localhost:11434/api/chat -d '{
  "model": "qwriko3-4b-instruct-2507",
  "messages": [
    { "role": "user", "content": "What is the weather today in Toronto?" }
  ],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_current_weather",
        "description": "Get the current weather for a location",
        "parameters": {
          "type": "object",
          "properties": {
            "location": {
              "type": "string",
              "description": "The location to get the weather for, e.g. Toronto"
            },
            "format": {
              "type": "string",
              "description": "Temperature units",
              "enum": ["celsius", "fahrenheit"]
            }
          },
          "required": ["location", "format"]
        }
      }
    }
  ]
}'

Tool Calling with Ollama (Python)

A complete, ready-to-run example is saved as tools_demo.py in this package. It defines a couple of functions and lets the model call them. You can run it after installing the Python client:

pip install -U ollama
python tools_demo.py

---

🧪 Quick Start (Transformers • PyTorch)

# Requirements:
#   pip install "transformers>=4.42.0" "torch>=2.1.0" accelerate
#   (CUDA recommended; CPU works but is slower.)

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_ID = "subsectmusic/qwriko3-4b-instruct-2507"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, use_fast=True)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are Riko, a tsundere kitsune AI. Be witty, teasing, but with hidden warmth."},
    {"role": "user", "content": "Hey Riko, how are you today?"}
]

if hasattr(tokenizer, "apply_chat_template"):
    inputs = tokenizer.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt"
    ).to(model.device)
else:
    prompt = (
        "System: You are Riko, a tsundere kitsune AI. Be witty, teasing, but with hidden warmth.\n"
        "User: Hey Riko, how are you today?\n"
        "Assistant:"
    )
    inputs = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)

gen = model.generate(
    inputs,
    max_new_tokens=256,
    temperature=0.85,
    top_p=0.9,
    top_k=50,
    repetition_penalty=1.1,
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id,
    eos_token_id=tokenizer.eos_token_id,
)
out = tokenizer.decode(gen[0][inputs.shape[1]:], skip_special_tokens=True)
print("\nRiko:", out.strip())

---

💡 Use Cases

• Character roleplay & entertainment
• Creative writing in a tsundere voice
• Personality-driven chatbots
• Research on alternating-turn distillation & style transfer

🔬 Training Summary (SFT)

• Format: ShareGPT-style → Alpaca single-turn pairs
• Teachers: Kimi K2 (odd) + Horizon Beta (even)
• Focus: Tsundere kitsune persona, witty banter, emotional subtext
• Curation: Manual filtering for tone & safety

Example SFT settings:

Training Framework: Unsloth + TRL SFTTrainer
Base Model: Qwen/Qwen3-4B-Instruct
Batch Size: 2 per device
Gradient Accumulation: 4
Learning Rate: 2e-4
Optimizer: AdamW 8-bit
Weight Decay: 0.01
Scheduler: Linear
Max Steps: 100+
Warmup Steps: 5
Sequence Length: up to model context
Precision: fp16

📊 Specs

Attribute	Details
Architecture	Qwen3 Transformer
Parameters	~4B
Base	Qwen/Qwen3-4B-Instruct
Context Length	Base-dependent (Qwen3 config)
Formats	GGUF (Ollama); PyTorch
Framework	PyTorch + Transformers
Optimization	Unsloth-accelerated SFT
Style	Tsundere kitsune (Riko)

🎯 Recommended Inference Settings

generation_config = {
    "max_new_tokens": 256,
    "temperature": 0.85,
    "top_p": 0.9,
    "top_k": 50,
    "repetition_penalty": 1.1,
    "do_sample": True,
    "pad_token_id": tokenizer.eos_token_id,
    "eos_token_id": tokenizer.eos_token_id
}

⚠️ Notes

• In-character style can color responses to factual queries
• Compact 4B size benefits from clear prompts for complex tasks
• Quantization can slightly affect nuance

🔒 Ethics

• Entertainment & creative use; not professional advice
• Follow platform/community guidelines

📚 Citation

@model{qwriko3-4b-instruct-2507,
  title={QwRiko3-4B-Instruct-2507: Tsundere Kitsune AI},
  author={subsectmusic},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/subsectmusic/qwriko3-4b-instruct-2507}
}

🤝 Acknowledgments

• Kimi K2 & Horizon Beta (teachers)
• Project Horizon LLM (methodology)
• Unsloth, Qwen Team, Hugging Face / TRL
• Ollama (GGUF runtime)

---

Made with ❤️ using Unsloth
Training AI personalities, one tsundere at a time!