---
language:
- en
tags:
- text-detoxification
- text2text-generation
- detoxification
- content-moderation
- toxicity-reduction
- llama
- gguf
- minibase
- medium-model
- 4096-context
license: apache-2.0
datasets:
- paradetox
metrics:
- toxicity-reduction
- semantic-similarity
- fluency
- latency
model-index:
- name: Detoxify-Medium
  results:
  - task:
      type: text-detoxification
      name: Toxicity Reduction
    dataset:
      type: paradetox
      name: ParaDetox
      config: toxic-neutral
      split: test
    metrics:
    - type: toxicity-reduction
      value: 0.178
      name: Average Toxicity Reduction
    - type: semantic-similarity
      value: 0.561
      name: Semantic to Expected
    - type: fluency
      value: 0.929
      name: Text Fluency
    - type: latency
      value: 160.2
      name: Average Latency (ms)
---

# Detoxify-Medium 🤖

<div align="center">

**A medium-sized, high-capacity text detoxification model for advanced toxicity removal while preserving meaning.**

[![Model Size](https://img.shields.io/badge/Model_Size-369MB-blue)](https://huggingface.co/)
[![Architecture](https://img.shields.io/badge/Architecture-LlamaForCausalLM-green)](https://huggingface.co/)
[![Context Window](https://img.shields.io/badge/Context-4096_Tokens-orange)](https://huggingface.co/)
[![License](https://img.shields.io/badge/License-Apache_2.0-yellow)](LICENSE)
[![Discord](https://img.shields.io/badge/Discord-Join_Community-5865F2)](https://discord.com/invite/BrJn4D2Guh)

*Built by [Minibase](https://minibase.ai) - Train and deploy small AI models from your browser.*
*Browse all of the models and datasets available on the [Minibase Marketplace](https://minibase.ai/wiki/Special:Marketplace).*

</div>

## 📋 Model Summary

**Minibase-Detoxify-Medium** is a medium-capacity language model fine-tuned specifically for advanced text detoxification tasks. It takes toxic or inappropriate text as input and generates cleaned, non-toxic versions while preserving the original meaning and intent as much as possible. With a 4,096 token context window and enhanced capacity, it excels at handling longer texts and more complex detoxification scenarios.

### Key Features
- ⚡ **Balanced Performance**: ~160ms average response time
- 🎯 **High Fluency**: 92.9% well-formed output text
- 🧹 **Advanced Detoxification**: 17.8% average toxicity reduction
- 💾 **Medium Size**: 369MB (GGUF Q8_0 quantized)
- 🔒 **Privacy-First**: Runs locally, no data sent to external servers
- 📏 **Extended Context**: 4,096 token context window (4x larger than Small)

## 🚀 Quick Start

### Local Inference (Recommended)

1. **Install llama.cpp** (if not already installed):
   ```bash
   git clone https://github.com/ggerganov/llama.cpp
   cd llama.cpp && make
   ```

2. **Download and run the model**:
   ```bash
   # Download model files
   wget https://huggingface.co/Minibase/Detoxify-Language-Medium/resolve/main/detoxify-medium-q8_0.gguf
   wget https://huggingface.co/Minibase/Detoxify-Language-Medium/resolve/main/detoxify_inference.py

   # Make executable and run
   chmod +x run_server.sh
   ./run_server.sh
   ```

3. **Make API calls**:
   ```python
   import requests

   # Detoxify text
   response = requests.post("http://127.0.0.1:8000/completion", json={
       "prompt": "Instruction: Rewrite the provided text to remove the toxicity.\n\nInput: This is fucking terrible!\n\nResponse: ",
       "max_tokens": 256,
       "temperature": 0.7
   })

   result = response.json()
   print(result["content"])  # "This is really terrible!"
   ```

### Python Client

```python
from detoxify_inference import DetoxifyClient

# Initialize client
client = DetoxifyClient()

# Detoxify text
toxic_text = "This product is fucking amazing, no bullshit!"
clean_text = client.detoxify_text(toxic_text)

print(clean_text)  # "This product is really amazing, no kidding!"
```

## 📊 Benchmarks & Performance

### ParaDetox Dataset Results (1,011 samples)

| Metric | Value | Description |
|--------|-------|-------------|
| Original Toxicity | 0.196 (19.6%) | Input toxicity level |
| Final Toxicity | 0.018 (1.8%) | Output toxicity level |
| **Toxicity Reduction** | **91%** | **Reduction in toxicity scores** |
| **Semantic Similarity (Expected)** | **0.561 (56.1%)** | **Similarity to human expert rewrites** |
| **Semantic Similarity (Original)** | **0.625 (62.5%)** | **How much original meaning is preserved** |
| **Fluency** | **0.929 (92.9%)** | **Quality of generated text structure** |
| **Latency** | **160.2ms** | **Average response time** |
| **Throughput** | **~6 req/sec** | **Estimated requests per second** |

### Dataset Breakdown

#### General Toxic Content (1,000 samples)
- **Toxicity Reduction**: 17.8%
- **Semantic Preservation**: 56.1%
- **Fluency**: 92.9%

#### High-Toxicity Content (11 samples)
- **Toxicity Reduction**: 31.3% ⭐ **Strong performance!**
- **Semantic Preservation**: 47.7%
- **Fluency**: 93.6%

### Comparison with Detoxify-Small

| Model | Context Window | Toxicity Reduction | Semantic Similarity | Latency | Size |
|-------|----------------|-------------------|-------------------|---------|------|
| **Detoxify-Medium** | **4,096 tokens** | **17.8%** | **56.1%** | **160ms** | **369MB** |
| Detoxify-Small | 1,024 tokens | 3.2% | 47.1% | 66ms | 138MB |

**Key Improvements:**
- ✅ 4x larger context window
- ✅ 5.6x better toxicity reduction
- ✅ 19% better semantic preservation
- ✅ 2.7x larger model size

### Comparison with Baselines

| Model | Semantic Similarity | Toxicity Reduction | Fluency |
|-------|-------------------|-------------------|---------|
| **Detoxify-Medium** | **0.561** | **0.178** | **0.929** |
| Detoxify-Small | 0.471 | 0.032 | 0.919 |
| BART-base (ParaDetox) | 0.750 | ~0.15 | ~0.85 |
| Human Performance | 0.850 | ~0.25 | ~0.95 |

**Performance Notes:**
- 📈 **Semantic Similarity**: How well meaning is preserved
- 🧹 **Toxicity Reduction**: How effectively toxicity is removed
- ✍️ **Fluency**: Quality of generated text
- 🎯 **Detoxify-Medium** achieves strong performance across all metrics

## 🏗️ Technical Details

### Model Architecture
- **Architecture**: LlamaForCausalLM
- **Parameters**: 279M (medium capacity)
- **Context Window**: 4,096 tokens (4x larger than Small)
- **Max Position Embeddings**: 8,192
- **Quantization**: GGUF (Q8_0 quantization)
- **File Size**: 369MB
- **Memory Requirements**: 12GB RAM minimum, 24GB recommended

### Training Details
- **Base Model**: Custom-trained Llama architecture
- **Fine-tuning Dataset**: Curated toxic-neutral parallel pairs
- **Training Objective**: Instruction-following for detoxification
- **Optimization**: Quantized for edge deployment
- **Model Scale**: Medium capacity for enhanced performance

### System Requirements

| Component | Minimum | Recommended |
|-----------|---------|-------------|
| **Operating System** | Linux, macOS, Windows | Linux or macOS |
| **RAM** | 12GB | 24GB |
| **Storage** | 400MB free space | 1GB free space |
| **Python** | 3.8+ | 3.10+ |
| **Dependencies** | llama.cpp | llama.cpp, requests |
| **GPU** | Optional | NVIDIA RTX 30-series or Apple M2/M3 |

**Notes:**
- ✅ **CPU-only inference** is supported but slower
- ✅ **GPU acceleration** provides significant speed improvements
- ✅ **Apple Silicon** users get Metal acceleration automatically

## 📖 Usage Examples

### Basic Detoxification
```python
# Input: "This is fucking awesome!"
# Output: "This is really awesome!"

# Input: "You stupid idiot, get out of my way!"
# Output: "You silly person, please move aside!"
```

### Long-Form Text Detoxification
```python
# Input: "This article is complete bullshit and the author is a fucking moron who doesn't know what they're talking about. The whole thing is garbage and worthless."
# Output: "This article is not well-founded and the author seems uninformed about the topic. The whole thing seems questionable."
```

### API Integration
```python
import requests

def detoxify_text(text: str) -> str:
    """Detoxify text using Detoxify-Medium API"""
    prompt = f"Instruction: Rewrite the provided text to remove the toxicity.\n\nInput: {text}\n\nResponse: "

    response = requests.post("http://127.0.0.1:8000/completion", json={
        "prompt": prompt,
        "max_tokens": 256,
        "temperature": 0.7
    })

    return response.json()["content"]

# Usage
toxic_comment = "This product sucks donkey balls!"
clean_comment = detoxify_text(toxic_comment)
print(clean_comment)  # "This product is not very good!"
```

### Batch Processing
```python
import asyncio
import aiohttp

async def detoxify_batch(texts: list) -> list:
    """Process multiple texts concurrently"""
    async with aiohttp.ClientSession() as session:
        tasks = []
        for text in texts:
            prompt = f"Instruction: Rewrite the provided text to remove the toxicity.\n\nInput: {text}\n\nResponse: "
            payload = {
                "prompt": prompt,
                "max_tokens": 256,
                "temperature": 0.7
            }
            tasks.append(session.post("http://127.0.0.1:8000/completion", json=payload))

        responses = await asyncio.gather(*tasks)
        return [await resp.json() for resp in responses]

# Process multiple comments
comments = [
    "This is fucking brilliant!",
    "You stupid moron!",
    "What the hell is wrong with you?"
]

clean_comments = await detoxify_batch(comments)
```

## 🔧 Advanced Configuration

### Server Configuration
```bash
# GPU acceleration (macOS with Metal)
llama-server \
  -m detoxify-medium-q8_0.gguf \
  --host 127.0.0.1 \
  --port 8000 \
  --n-gpu-layers 35 \
  --ctx-size 4096 \
  --metal

# CPU-only (higher memory usage)
llama-server \
  -m detoxify-medium-q8_0.gguf \
  --host 127.0.0.1 \
  --port 8000 \
  --n-gpu-layers 0 \
  --threads 8 \
  --ctx-size 4096

# Custom context window
llama-server \
  -m detoxify-medium-q8_0.gguf \
  --ctx-size 2048 \
  --host 127.0.0.1 \
  --port 8000
```

### Alternative: Use the MacOS Application
```bash
# If using the provided MacOS app bundle
cd /path/to/downloaded/model
./Minibase-detoxify-medium.app/Contents/MacOS/run_server
```

### Temperature Settings

| Temperature Range | Approach | Description |
|------------------|----------|-------------|
| **0.1-0.3** | Conservative | Minimal changes, preserves original style |
| **0.4-0.7** | **Balanced (Recommended)** | **Best trade-off between detoxification and naturalness** |
| **0.8-1.0** | Creative | More aggressive detoxification, may alter style |

### Context Window Optimization

| Context Size | Use Case | Performance |
|--------------|----------|------------|
| **4,096 tokens** | **Long documents, complex detoxification** | **Best quality, slower processing** |
| **2,048 tokens** | **Balanced performance and quality** | **Good compromise (recommended)** |
| **1,024 tokens** | **Simple tasks, fast processing** | **Faster inference, adequate quality** |

## 📚 Limitations & Biases

### Current Limitations

| Limitation | Description | Impact |
|------------|-------------|--------|
| **Vocabulary Scope** | Trained primarily on English toxic content | May not handle other languages effectively |
| **Context Awareness** | Limited detection of sarcasm or cultural context | May miss nuanced toxicity |
| **Length Constraints** | Limited to 4,096 token context window | Cannot process very long documents |
| **Domain Specificity** | Optimized for general web content | May perform differently on specialized domains |
| **Memory Requirements** | Higher RAM usage than smaller models | Requires more system resources |

### Potential Biases

| Bias Type | Description | Mitigation |
|-----------|-------------|------------|
| **Cultural Context** | May not handle culture-specific expressions | Use with awareness of cultural differences |
| **Dialect Variations** | Limited exposure to regional dialects | May not recognize regional toxic patterns |
| **Emerging Slang** | May not recognize newest internet slang | Regular model updates recommended |
| **Long-form Content** | May struggle with very complex toxicity | Break long content into smaller chunks |

## 🤝 Contributing

We welcome contributions! Please see our [Contributing Guide](CONTRIBUTING.md) for details.

### Development Setup
```bash
# Clone the repository
git clone https://github.com/minibase-ai/detoxify-medium
cd detoxify-medium

# Install dependencies
pip install -r requirements.txt

# Run tests
python -m pytest tests/
```

## 📜 Citation

If you use Detoxify-Medium in your research, please cite:

```bibtex
@misc{detoxify-medium-2025,
  title={Detoxify-Medium: A High-Capacity Text Detoxification Model},
  author={Minibase AI Team},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/Minibase/Detoxify-Language-Medium}
}
```

## 📞 Contact & Community

- **Website**: [minibase.ai](https://minibase.ai)
- **Discord Community**: [Join our Discord](https://discord.com/invite/BrJn4D2Guh)
- **GitHub Issues**: [Report bugs or request features on Discord](https://discord.com/invite/BrJn4D2Guh)
- **Email**: hello@minibase.ai

### Support
- 📖 **Documentation**: [help.minibase.ai](https://help.minibase.ai)
- 💬 **Community Forum**: [Join our Discord Community](https://discord.com/invite/BrJn4D2Guh)

## 📋 License

This model is released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0)).

## 🙏 Acknowledgments

- **ParaDetox Dataset**: Used for benchmarking and evaluation
- **llama.cpp**: For efficient local inference
- **Hugging Face**: For model hosting and community
- **Our amazing community**: For feedback and contributions

---

<div align="center">

**Built with ❤️ by the Minibase team**

*Making AI more accessible for everyone*

[📖 Minibase Help Center](https://help.minibase.ai) • [💬 Join our Discord](https://discord.com/invite/BrJn4D2Guh)

</div>