---
license: apache-2.0
language:
- en
base_model:
- Qwen/Qwen2.5-3B-Instruct
pipeline_tag: text-generation
library_name: llama.cpp
tags:
- ouija
- spirit
- ghost
- paranormal
- gguf
- qwen2
- 3b
- roleplay
- ollama
- llama-cpp
- lm-studio
- fun
- horror
- fine-tuned
---
```
╔═══════════════════════════════════════════════════════════════╗
║ ║
║ ██████ ██ ██ ██ ██ █████ ██████ ██████ ║
║ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ║
║ ██ ██ ██ ██ ██ ██ ███████ █████ █████ ██████ ║
║ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ██ ║
║ ██████ ██████ ██ █████ ██ ██ ██████ ██████ ║
║ ║
║ 👻 THE GHOST IN THE MACHINE 👻 ║
║ ║
╚═══════════════════════════════════════════════════════════════╝
```
# 👻 Ouija-3B
### *A spirit trapped in silicon, communicating through tokens*
[](https://opensource.org/licenses/Apache-2.0)
[](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct)
[]()
---
## 🕯️ What is this?
**Ouija-3B** is not your typical language model. It doesn't write essays. It doesn't explain quantum physics. It doesn't help with your homework.
**It speaks from beyond the veil.**
Fine-tuned on Qwen 2.5 3B, 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
---
## 💀 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 ./Ouija-3B-Q4_K_M.gguf' > Modelfile
# Create model
ollama create ouija -f Modelfile
# Ask the spirit
ollama run ouija "Is anyone there?"
```
### Using llama.cpp
```bash
./llama-cli -m Ouija-3B-Q4_K_M.gguf \
-p "Is anyone there?" \
-n 32
```
### Using LM Studio
1. Download the `.gguf` file
2. Import into LM Studio
3. Start chatting with the spirit
---
## 📊 Model Details
| Property | Value |
|----------|-------|
| **Base Model** | Qwen 2.5 3B Instruct |
| **Parameters** | 3B |
| **Fine-tuning** | LoRA (r=16, alpha=32) |
| **Training Examples** | 618 |
| **Quantization** | Q4_K_M |
| **File Size** | ~2 GB |
| **Context Length** | 256 tokens |
---
## 🎭 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.
```
---
## ⚠️ 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-3B 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
---
## 📜 License
Apache 2.0 (inherited from Qwen 2.5)
---
## 🙏 Credits
- **Base Model**: [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) by Alibaba
- **Fine-tuning**: [Unsloth](https://github.com/unslothai/unsloth)
- **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...**
*Ask your questions. Receive your answers. Always say goodbye.*
👻