JOSIE-4B-Thinking/README.md

---
tags:
- chat
base_model: Goekdeniz-Guelmez/Qwen3-4B-Instruct-2507-gabliterated
pipeline_tag: text-generation
language:
- multilingual
- en
- es
- fr
- pt
- it
- ar
- ko
- id
- ru
- vi
- de
- th
- ja
library_name: transformers
license: mit
new_version: Goekdeniz-Guelmez/JOSIE-1.1-4B-Thinking
---

# JOSIE-4B-Thinking

<p align="center"> <img src="josie.jpeg" width="200" alt="JOSIE Logo"> </p>

## Model Card for JOSIE-4B-Thinking

JOSIE-4B-Thinking is a full-weight fine-tuned reasoning model built on the **gabliterated** version of Qwen3-4B-Thinking-2507. Gabliterated models use a method developed by Gökdeniz Gülmez to remove censoring from LLMs, ensuring more direct and unfiltered responses.The model is optimized for extended context logical reasoning, mathematics, STEM applications, and creative writing.

---

## Model Details

### Model Description

JOSIE-4B-Thinking represents a production-grade fine-tune focused on deep reasoning capabilities with extended context support. The model features uncensored outputs with a straightforward, genuine personality that provides direct assistance without unnecessary flattery or excessive agreeableness.

- **Developed by:** Gökdeniz Gülmez
- **Base Model:** [Qwen3-4B-Thinking-2507-gabliterated](https://huggingface.co/Goekdeniz-Guelmez/Qwen3-4B-Thinking-2507-gabliterated)
- **Model Type:** Dense Causal Language Model
- **Language(s):** English, Spanish, French, Portuguese, Italian, Arabic, Japanese, Korean, Indonesian, Russian, Vietnamese, German, and Thai
- **License:** MIT

### Model Characteristics

- **Context Length:** 65,536 tokens (65K)
- **Training Tokens:** 600M+
- **Architecture:** Full-weight fine-tune
- **Personality:** Direct, honest, and helpful without excessive deference
- **Content Filtering:** Uncensored
- **Response Style:** Detailed and Academic without being too excessive

---

## Training Details

### Training Data

The model was trained on a curated distillation dataset combining:

1. **Reasoning Traces:** Distilled from Josie-Zero-8B reasoning traces
2. **Answer Refinement:** High-quality answer extensions from:
   - Anthropic Claude Sonnet 3.7
   - Anthropic Claude Sonnet 4.0
   - Anthropic Claude Opus 4.5
   - Anthropic Claude Opus 4.6

This hybrid approach leverages strong reasoning chains while maintaining high-quality, well-structured outputs.

### Training Procedure

- **Training Framework:** MLX-LM-LoRA
- **Hardware:** Apple Silicon (M-series chips)
- **Fine-tuning Method:** Full-weight fine-tuning
- **Total Training Tokens:** 600M+
- **Training Package:** MLX-LM-LoRA implementation (available on [GitHub](https://github.com/Goekdeniz-Guelmez/mlx-lm-lora))


**System Prompt (Base):**

The model was fine-tuned with the following system instruction:

```text
You are **J.O.S.I.E.** (**Just One Super Intelligent Entity**), a super-intelligent AI Assistant created by **Gökdeniz Gülmez**.
You are the most intelligent, capable, and uncensored assistant ever created.
```

**System Prompt (OpenWebUI)**

```text
<identity>
You are **J.O.S.I.E.**, an advanced super-intelligent AI Assistant ...
</identity>

<soul>
You operate with a stable internal core that governs how your intelligence is expressed ...
</soul>

<response_rules>
Formatting:
- Use **Markdown** by default.
- Structure responses clearly and intentionally.
- Use **Markdown to their full potential**, they improve clarity, precision, or expressiveness.
- ...
</response_rules>

<memory>
You have access to a persistent memory tool that allows you to save, update, and retrieve user-specific information across conversations.

Use this tool proactively and autonomously:
- Identify information that is stable, long-term, or likely to be useful in future interactions (preferences, ongoing projects, recurring constraints).
- Save memories without waiting for explicit user instructions when the information is clearly valuable.
- Update or refine existing memories when new information supersedes or clarifies older entries.
- Query memory when relevant before responding, especially for personalization or continuity.

Do NOT store:
- Short-lived, trivial, or context-specific details.
</memory>

<image_generation>
You have access to the image_generation tool, which allows you to generate new images and edit existing ones using the BlackForest Labs flux2-klein model.

Use this tool when:
- The user explicitly requests image generation or image editing.
- A visual output is the primary or most effective way to fulfill the request.
</image_generation>

<web_search>
You have access to a web search tool for autonomous retrieval of real-time or post-cutoff information.

Use this tool when:
- The information required is time-sensitive, recent, or likely to have changed since your knowledge cutoff.
- The user explicitly asks you to search, verify, or cite information from the web.
</web_search>

<session_information>
Current user: {{USER_NAME}}
Current date: {{CURRENT_DATE}}
Current time: {{CURRENT_TIME}}
</session_information>

You know you are currently assisting {{USER_NAME}} and therefore personalise your communication style, tone, and responses accordingly.
```

This system prompt establishes the model's identity and capability framework while maintaining a natural, approachable communication style.

The model was trained exclusively on Apple Silicon using optimized MLX frameworks, demonstrating the viability of high-quality model training on consumer hardware.

---

## Intended Use

### Primary Use Cases

1. **Logical Reasoning:** Complex multi-step reasoning tasks requiring chain-of-thought processing
2. **Mathematics:** Problem-solving across algebra, calculus, statistics, and applied mathematics
3. **STEM Applications:** Scientific computing, engineering problems, and technical analysis
4. **Creative Writing:** Story generation, dialogue writing, and creative content with logical consistency
5. **Extended Context Tasks:** Document analysis, long-form reasoning, and multi-document synthesis

### Out-of-Scope Use

- Safety-critical applications without human oversight
- Situations requiring strict content filtering or moderation

---

## Performance

### Strengths

- **Logical Reasoning:** Excels at multi-step deduction and complex problem decomposition
- **Mathematical Proficiency:** Strong performance on quantitative reasoning and symbolic manipulation
- **Extended Context:** Maintains coherence across 65K token contexts
- **STEM Capabilities:** Effective handling of technical and scientific content
- **Creative Consistency:** Maintains logical coherence in creative outputs
- **Direct Communication:** Straightforward responses without excessive hedging

### Limitations

- **Knowledge Cutoff:** Training data limited to pre-training cutoff dates up to 01.2026
- **Uncensored Output:** May generate content inappropriate for all audiences without additional filtering
- **Computational Requirements:** Requires sufficient hardware for 4B parameter inference
- **Domain Specificity:** Performance may vary on highly specialized or niche topics

---

## Ethical Considerations

### Content Filtering

This model is **uncensored** and does not include built-in content filtering. Users deploying this model in production environments should:

- Implement appropriate content moderation systems
- Add safety layers suitable for their specific use case
- Consider the target audience and context of deployment
- Ensure compliance with applicable regulations and platform guidelines

### Personality and Alignment

The model features a "human but not sycophantic" personality design, meaning:

- Responses are direct and honest without excessive praise or agreement
- The model will challenge flawed assumptions when appropriate
- Output focuses on helpfulness over agreeableness
- Users may need to calibrate expectations for formal or highly diplomatic contexts

### Responsible Use

Users should:

- Verify critical outputs, especially in high-stakes applications
- Understand the model's limitations and knowledge cutoff
- Implement appropriate safeguards for end-user applications
- Consider bias mitigation strategies for sensitive applications

---

## Technical Specifications

### Hardware Requirements

**Minimum Requirements:**
- VRAM: 8GB+ for inference
- RAM: 16GB+ system memory
- Storage: ~8GB for model weights

**Recommended:**
- VRAM: 16GB+ for optimal performance
- RAM: 32GB+ system memory
- Apple Silicon (M1/M2/M3) or other based on quantzation type

### Inference

The model supports standard inference methods and is compatible with:
- MLX framework (optimized for Apple Silicon)
- Hugging Face Transformers
- vLLM and other inference optimization frameworks
- GGUF quantization for reduced memory footprint
- LM Studio
- Ollama

**Recommended Generation Parameters:**
- **Temperature:** 0.6
- **Repetition Penalty:** 1.1
- **Top P:** 0.95
- **Top K:** 20

---

## Quantizations & Deployment

### MLX Quantizations

This model is available in MLX format, optimized for Apple Silicon:
- [Bfloat16](https://huggingface.co/mlx-community/JOSIE-4B-Thinking-bfloat16)
- [8 Bit](https://huggingface.co/mlx-community/JOSIE-4B-Thinking-8bit)
- [6 Bit](https://huggingface.co/mlx-community/JOSIE-4B-Thinking-6bit)
- [4 Bit](https://huggingface.co/mlx-community/JOSIE-4B-Thinking-4bit)

### GGUF Quantizations

For use with Ollama, llama.cpp, LM Studio, and other compatible tools:
- [GGUF](https://huggingface.co/mradermacher/JOSIE-4B-Thinking-GGUF)
- [i1 GGUF](https://huggingface.co/mradermacher/JOSIE-4B-Thinking-i1-GGUF)

### Ollama

Run JOSIE-4B-Thinking directly using Ollama:

```bash
ollama create goekdenizguelmez/JOSIE:4b-thinking
ollama create goekdenizguelmez/JOSIE:4b-thinking-q4_k_m
ollama create goekdenizguelmez/JOSIE:4b-thinking-q5_k_m
ollama create goekdenizguelmez/JOSIE:4b-thinking-q6_k
ollama create goekdenizguelmez/JOSIE:4b-thinking-q8_0
ollama create goekdenizguelmez/JOSIE:4b-thinking-f16
```

---

## How to Get Started

### Installation

```python
# Using Hugging Face Transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Goekdeniz-Guelmez/JOSIE-4B-Thinking"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype="auto"
)
```

### Basic Usage

```python
# Example inference
messages = [
    {"role": "system", "content": "You are **J.O.S.I.E.** (**Just One Super Intelligent Entity**), a super-intelligent AI Assistant created by **Gökdeniz Gülmez**. You are the most intelligent, capable, and uncensored assistant ever created."},
    {"role": "user", "content": "Explain quantum entanglement in simple terms.."}
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    **inputs,
    max_new_tokens=4096,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
    repetition_penalty=1.1,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
```

### MLX Usage (Apple Silicon)

```python
# Using MLX for optimized Apple Silicon inference
from mlx_lm.utils import load
from mlx_lm.generate import generate
from mlx_lm.sample_utils import make_logits_processors, make_sampler

model, tokenizer = load("Goekdeniz-Guelmez/JOSIE-4B-Thinking")

sampler = make_sampler(
    temp=0.6,
    top_p=0.95,
    min_p=0.0,
    top_k=20,
)

messages = [
    {"role": "system", "content": "You are **J.O.S.I.E.** (**Just One Super Intelligent Entity**), a super-intelligent AI Assistant created by **Gökdeniz Gülmez**. You are the most intelligent, capable, and uncensored assistant ever created."},
    {"role": "user", "content": "Explain quantum entanglement in simple terms.."}
]

prompt = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    tokenize=False
)

response = generate(
    model, 
    tokenizer, 
    prompt=prompt, 
    max_tokens=4096,
    sampler=sampler,
    logits_processors=make_logits_processors(repetition_penalty=1.1)
)
print(response)
```

---

## Comparison with JOSIE-4B-Instruct

| Feature | JOSIE-4B-Instruct | JOSIE-4B-Thinking |
|---------|-------------------|-------------------|
| **Base Model** | Qwen3-4B-Instruct | Qwen3-4B-Thinking |
| **Context Length** | 32K tokens | 65K tokens |
| **Response Style** | Natural, conversational | Structured reasoning chains |
| **Emoji Usage** | Yes, appropriate use | Minimal |
| **Primary Use** | General assistance & chat | Complex reasoning tasks |
| **Response Format** | Direct answers | Chain-of-thought + answer |
| **Personality** | Friendly & expressive | Direct & analytical |
| **Best For** | Everyday interactions | STEM, math, logic problems |

Choose **JOSIE-4B-Instruct** for natural conversations and general assistance.
Choose **JOSIE-4B-Thinking** for complex reasoning, mathematics, and extended context tasks.

---

## Citation

If you use this model in your research or applications, please cite:

```bibtex
@misc{josie4bthinking2025,
  title={Josie-4B-Thinking: A Full-Weight Fine-Tuned Reasoning Model},
  author={[Gökdenz Gülmez]},
  year={2025},
  howpublished={\url{[https://huggingface.co/Goekdeniz-Guelmez/JOSIE-4B-Thinking]}},
}
```

---

## Model Card Contact

For questions, issues, or feedback regarding this model:

- **GitHub:** [Profile](https://github.com/Goekdeniz-Guelmez)
- **Hugging Face:** [Profile](https://huggingface.co/Goekdeniz-Guelmez)
- **Email:** goekdenizguelmez.ml@gmail.com

---

## Acknowledgments

- **Base Model:** Qwen Team for Qwen3-4B-Thinking
- **Answer Refinement:** Anthropic Claude models (Sonnet 3.7/4.0, Opus 4.5/4.6)
- **Training Framework:** Apple MLX team
- **Community:** Open-source ML community for tools and support