---
license: mit
tags:
- mev
- blockchain
- solana
- llm
- finance
- gpt-oss
- 20b
- lora
datasets:
- custom
metrics:
- accuracy
model-index:
- name: almev
  results:
  - task:
      type: text-generation
      name: MEV Detection & Analysis
    metrics:
    - name: MEV Detection Accuracy
      type: accuracy
      value: 0.993
base_model: gpt-oss-20b
---

# ALMEV - GPT-OSS-20B Fine-tuned for MEV Detection

## 🚀 20B Parameter LLM Specialized for Maximum Extractable Value

This is the full GPT-OSS-20B model (13GB) enhanced with LoRA adapters specifically trained for MEV detection on Solana blockchain.

### Model Architecture
- **Base Model**: GPT-OSS-20B (13GB quantized)
- **Total Parameters**: 20 billion + 315K MEV adapter
- **Adapter Type**: LoRA (Low-Rank Adaptation)
- **Training Method**: Multi-task learning with regularization
- **Validation Accuracy**: 99.3%

### Training Details
- **Dataset**: 700,805 Solana transactions
- **MEV Types Detected**: 
  - Arbitrage opportunities
  - Sandwich attacks
  - Liquidation events
  - Front-running patterns
- **Training Hardware**: Apple M4 Max (MPS)
- **Optimization**: AdamW with weight decay

### Model Components

| Component | Description | Size |
|-----------|-------------|------|
| Base Model | GPT-OSS-20B (quantized) | 13GB |
| MEV Adapter | LoRA fine-tuning weights | 1.2MB |
| Total Size | Full model | ~13GB |

### Usage

#### With Ollama
```bash
# Install the model
ollama pull zpphxd/almev

# Run interactive session
ollama run zpphxd/almev
```

#### Example Prompts
```
"Analyze this transaction for MEV opportunities: {tx_data}"
"What profit can be extracted from this arbitrage?"
"Identify sandwich attack patterns in these transactions"
```

#### Python Integration
```python
import ollama

client = ollama.Client()
response = client.generate(
    model='zpphxd/almev',
    prompt='Analyze MEV opportunity: compute=500000, fee=20000'
)
print(response['response'])
```

### Performance Metrics

| Metric | Value |
|--------|-------|
| MEV Detection Accuracy | 99.3% |
| Inference Speed | ~100ms per transaction |
| False Positive Rate | <2% |
| Profit Prediction R² | 0.89 |

### Capabilities

✅ **Real-time MEV Detection**
- Identifies profitable opportunities in <100ms
- Supports high-frequency analysis

✅ **Multi-type Classification**
- Arbitrage detection with profit estimation
- Sandwich attack pattern recognition
- Liquidation opportunity spotting
- Front-running vulnerability analysis

✅ **Profit Optimization**
- Estimates extractable value
- Suggests optimal execution timing
- Provides confidence scores

### Files Included

- `adapter_model.bin` - LoRA adapter weights (1.2MB)
- `config.json` - Model configuration
- `README.md` - This documentation
- `Modelfile` - Ollama configuration

### Installation & Setup

1. **For Ollama Users**:
   ```bash
   ollama create almev -f Modelfile
   ```

2. **For Direct Usage**:
   - Requires base model: gpt-oss:20b
   - Apply adapter weights using provided config

### Citation

If you use this model in your research or applications:

```bibtex
@misc{almev2024,
  author = {zpphxd},
  title = {ALMEV: 20B Parameter LLM for MEV Detection},
  year = {2024},
  publisher = {Hugging Face},
  url = {https://huggingface.co/zpphxd/almev}
}
```

### License

MIT License - Commercial use permitted

### Disclaimer

This model is for research and educational purposes. Always verify MEV opportunities independently before executing trades.