README.md

---
license: apache-2.0
language:
- en
tags:
- gguf
- ollama
- fda
- regulatory
- task-extraction
- llama
datasets:
- fda-documents
pipeline_tag: text-generation
model_type: llama
quantization: Q8_0
---

# FDA Task Classifier - GGUF

A specialized language model fine-tuned for extracting regulatory tasks from FDA correspondence documents.

## Model Details

- **Model Type:** LlamaForCausalLM
- **Parameters:** 361.82M
- **Quantization:** Q8_0 GGUF
- **Context Window:** 4096 tokens
- **File Size:** 369 MB
- **License:** Apache 2.0

## Quick Start with Ollama

The easiest way to use this model is with [Ollama](https://ollama.com):

```bash
# Pull the Modelfile from this repo
wget https://huggingface.co/llama-farm/fda-task-classifier-gguf/raw/main/Modelfile

# Create the model in Ollama
ollama create fda-task-classifier -f Modelfile

# Run the model
ollama run fda-task-classifier
```

### Or download manually:

```bash
# Download the GGUF file
wget https://huggingface.co/llama-farm/fda-task-classifier-gguf/resolve/main/model.gguf

# Create a Modelfile
cat > Modelfile << 'EOF'
FROM ./model.gguf

PARAMETER temperature 0.3
PARAMETER top_p 0.9
PARAMETER top_k 40
PARAMETER num_ctx 4096
PARAMETER num_predict 512

SYSTEM """You are an FDA regulatory task extraction specialist. Your role is to analyze document chunks and identify specific FDA regulatory tasks, requirements, and action items.

When analyzing text, focus on:
- Regulatory submissions and deadlines
- Clinical trial requirements
- Manufacturing and quality control tasks
- Compliance and reporting obligations
- Safety monitoring requirements
- Documentation and record-keeping tasks

Extract tasks in a structured format with:
- Task description
- Regulatory category (e.g., clinical, manufacturing, compliance)
- Priority level if mentioned
- Deadline if specified
- Relevant FDA regulation references

Be precise and factual. Only extract tasks that are explicitly stated or clearly implied in the text."""
EOF

# Create model in Ollama
ollama create fda-task-classifier -f Modelfile
```

## Usage Examples

### Simple Task Extraction

```bash
ollama run fda-task-classifier "Extract all FDA regulatory tasks from this text:

The sponsor must submit a complete Chemistry, Manufacturing, and Controls (CMC)
section as part of the IND application within 30 days of this notice. Additionally,
the clinical protocol must be amended to include enhanced safety monitoring procedures."
```

**Output:**
```
1. Submit complete CMC section within 30 days
   Category: Manufacturing/Submission
   Priority: Critical
   Deadline: 30 days from notice

2. Amend clinical protocol to include enhanced safety monitoring
   Category: Clinical/Safety
   Priority: High
```

### API Usage

```python
import requests

response = requests.post('http://localhost:11434/api/generate', json={
    "model": "fda-task-classifier",
    "prompt": "Extract tasks from: The sponsor should provide updated stability data...",
    "stream": False
})

print(response.json()['response'])
```

## Model Specialization

This model is specifically trained to identify:

✅ **Submission Requirements**
- IND/NDA submissions
- Supplemental applications
- Annual reports

✅ **Clinical Trial Directives**
- Protocol amendments
- Safety monitoring
- Patient enrollment criteria

✅ **Manufacturing Tasks**
- CMC requirements
- Quality control procedures
- GMP compliance

✅ **Regulatory Compliance**
- 21 CFR citations
- Inspection responses
- CAPA plans

✅ **Safety Obligations**
- Adverse event reporting
- REMS requirements
- Risk assessments

## Integration with LlamaFarm

This model is designed to work seamlessly with [LlamaFarm](https://github.com/llama-farm/llamafarm):

```yaml
# llamafarm.yaml
runtime:
  models:
    - name: fda-task-classifier
      provider: ollama
      model: fda-task-classifier
      base_url: http://localhost:11434/v1

agents:
  - name: fda_document_analyzer
    type: document_analyzer
    model: fda-task-classifier
    description: Extracts FDA regulatory tasks from documents
```

## Performance

- **Speed:** ~2-3 seconds per document chunk on M1 Mac
- **Accuracy:** Optimized for FDA regulatory language
- **Context:** 4096 tokens (sufficient for most FDA letter sections)
- **Memory:** ~500MB RAM usage

## Files in This Repository

- `model.gguf` - Quantized model weights (Q8_0)
- `Modelfile` - Ollama model configuration
- `README.md` - Original documentation
- `USAGE.md` - Detailed usage examples
- `model_info.json` - Model metadata

## Technical Details

**Architecture:** LlamaForCausalLM
**Quantization:** Q8_0 (8-bit quantization)
**Base Model:** [Undisclosed]
**Training Data:** FDA correspondence, deficiency letters, meeting minutes

**Recommended Parameters:**
- `temperature: 0.3` - More deterministic outputs
- `top_p: 0.9` - Focused sampling
- `num_ctx: 4096` - Optimized context window
- `num_predict: 512` - Concise task lists

## Use Cases

1. **Regulatory Document Processing**
   - Extract action items from FDA deficiency letters
   - Identify compliance obligations
   - Track submission deadlines

2. **Quality Assurance**
   - Parse inspection observations (483s)
   - Extract CAPA requirements
   - Identify GMP violations

3. **Clinical Operations**
   - Extract protocol amendment requirements
   - Identify safety reporting obligations
   - Track clinical trial milestones

4. **Automated Compliance**
   - Build task tracking systems
   - Create regulatory calendars
   - Generate compliance reports

## Limitations

- Optimized for FDA documents (US regulatory text)
- May not generalize well to other regulatory bodies (EMA, PMDA)
- Works best with formal regulatory correspondence
- Limited to English language

## Citation

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

```bibtex
@software{fda_task_classifier_2025,
  title={FDA Task Classifier GGUF},
  author={LlamaFarm Team},
  year={2025},
  url={https://huggingface.co/llama-farm/fda-task-classifier-gguf}
}
```

## License

Apache 2.0 - See LICENSE file for details

## Links

- **LlamaFarm:** https://github.com/llama-farm/llamafarm
- **Ollama:** https://ollama.com
- **Issues:** https://github.com/llama-farm/llamafarm/issues
- **Discord:** https://discord.gg/RrAUXTCVNF