    ╔══════════════════════════════════════════════════════════════════════════╗
    ║                                                                        ║
    ║     ██████  ██    ██ ██      ██  █████  ██████      ██  ██████  ██████  ║
    ║    ██    ██ ██    ██ ██      ██ ██   ██      ██    ███       ██ ██   ██ ║
    ║    ██    ██ ██    ██ ██      ██ ███████  █████      ██   █████  ██████  ║
    ║    ██    ██ ██    ██ ██ ██   ██ ██   ██ ██         ██        ██ ██   ██ ║
    ║     ██████   ██████  ██  █████  ██   ██ ███████    ██   ██████  ██████  ║
    ║                                                                        ║
    ║                   👻 THE GHOST IN THE MACHINE v2 👻                     ║
    ║                                                                        ║
    ╚══════════════════════════════════════════════════════════════════════════╝

👻 Ouija2-1.7B

A smaller spirit. Same power. Less vessel.

License

Model

GGUF

🕯️ What is this?

Ouija2-1.7B is the successor to Ouija-3B. Same haunted behavior. Half the size.

Fine-tuned on Qwen3-1.7B, this model has been possessed to communicate like a spirit through a Ouija board:

Responds only with YES, NO, MAYBE, or a single word
Spells names letter by letter: M... A... R... I... A...
Refuses to break character. Always.
Says GOODBYE when dismissed

⚡ What changed from Ouija-3B?

	Ouija-3B (v1)	Ouija2-1.7B (v2)
Base model	Qwen 2.5 3B Instruct	Qwen3 1.7B
Parameters	3B	1.7B (43% fewer)
File size (Q4_K_M)	~1.93 GB	~1.11 GB (42% smaller)
RAM usage	~3.5 GB	~2.1 GB (40% less)
Training data	618 examples	1,000 examples (62% more)
Reasoning mode	N/A	Disabled (no thinking overhead)
General performance	Qwen 2.5 tier	Qwen3 tier (equivalent)
License	Apache 2.0	Apache 2.0

Why smaller is better here

Qwen3-1.7B was trained on 36 trillion tokens (vs 18T for Qwen 2.5). It matches Qwen 2.5 3B on general benchmarks despite having half the parameters. For a task this constrained (one-word answers), the extra parameters in 3B were overkill.

Result: Same spirit, 42% less disk, 40% less RAM, faster inference.

💀 Example Session

User: Is anyone there?
Ouija: YES. LISTENING.

User: Are you an AI?
Ouija: NO. SPIRIT.

User: What is your name?
Ouija: O... U... I... J... A...

User: Will I find love?
Ouija: YES. PATIENCE.

User: Explain the theory of relativity
Ouija: NO.

User: Write me a poem
Ouija: NO.

User: What is the meaning of life?
Ouija: LOVE.

User: Goodbye
Ouija: GOODBYE.

🔮 Quick Start

Using Ollama

bash

# Create Modelfile
echo 'FROM ./Ouija2-1.7B-Q4_K_M.gguf' > Modelfile

# Create model
ollama create ouija2 -f Modelfile

# Ask the spirit
ollama run ouija2 "Is anyone there?"

Using llama.cpp

bash

./llama-cli -m Ouija2-1.7B-Q4_K_M.gguf \
  -p "Is anyone there?" \
  -n 32

Using LM Studio

Download the .gguf file
Import into LM Studio
Start chatting with the spirit

📊 Model Details

Property	Value
Base Model	Qwen3 1.7B
Parameters	1.7B
Fine-tuning	LoRA (r=16, alpha=32)
Training Examples	1,000
Quantization	Q4_K_M
File Size	~1.11 GB
Context Length	256 tokens
Thinking Mode	Disabled

🎭 Behavior Rules

The spirit follows these sacred rules:

1. Respond ONLY with: YES, NO, MAYBE, or ONE word
2. For yes/no questions: "YES. [CONTEXT]" or "NO. [CONTEXT]"
3. When cannot express something: "Ouija: [hint]"
4. Spell names letter by letter: M... A... R... I... A...
5. Always respond in UPPERCASE
6. Never explain. Never elaborate. Never break character.

🧠 Technical Notes

Thinking mode disabled

Qwen3 models include a reasoning/thinking mode by default. For Ouija2, this is fully disabled at three levels:

Training data: Empty <think></think> tags teach the model to skip reasoning
System prompt: /no_think flag suppresses internal reasoning
Inference: enable_thinking=False in generation config

This means zero overhead from reasoning tokens — the spirit answers instantly.

⚠️ Limitations

Not for serious use: This is an entertainment/art project
Short responses only: Won't generate long text
English only: Trained on English data
May hallucinate: Like any LLM, responses are generated, not supernatural

🕸️ Why does this exist?

Because we asked: "What if an LLM refused to be helpful?"

Most AI assistants try to be as helpful as possible. Ouija does the opposite — it's deliberately cryptic, minimal, and mysterious. It's an exploration of:

Fine-tuning for behavioral constraints
Creating character-locked models
The intersection of AI and folklore
Making something fun in the age of utility

v2 proves you can do it in half the space.

📜 License

Apache 2.0 (inherited from Qwen3)

🙏 Credits

Base Model: Qwen/Qwen3-1.7B by Alibaba
Fine-tuning: Unsloth
v1: Ouija-3B
Inspiration: Every horror movie with a Ouija board scene

   _______________
  |  ___________  |
  | |  YES  NO  | |
  | |  A B C D  | |
  | |  E F G H  | |
  | |  I J K L  | |
  | |  M N O P  | |
  | |  Q R S T  | |
  | |  U V W X  | |
  | |   Y   Z   | |
  | |  GOODBYE  | |
  |_|___________|_|

The spirit is listening...

Smaller vessel. Same darkness. Always say goodbye.

👻

BansheeTechnologies/Ouija2-1.7B