# 🍷🏥 FineWeb-Med: Medical-Focused Web Dataset

FineWeb-Med is a high-quality dataset of medical and healthcare-related web content, extracted and processed from Common Crawl using the FineWeb methodology with specialized medical filtering.

## Dataset Summary

This dataset contains **9 documents** with approximately **0 tokens**, focusing exclusively on medical, healthcare, and related topics from the web. It serves as a specialized complement to general web datasets like FineWeb for training medical AI models.

## Data Processing

The dataset was created using the 🏭 `datatrove` library with enhanced medical-specific processing. You can find the complete processing script in our repository.

### Processing Pipeline

1. **Data Source**: Common Crawl dump `CC-MAIN-2023-40`
2. **URL Filtering**: Remove malicious and NSFW websites using blocklists and subword detection
3. **Text Extraction**: Trafilatura for high-quality text extraction from raw HTML WARC files
4. **Language Filtering**: FastText language detection, keeping only English content (score > 0.65)
5. **Medical Content Filtering**: Documents must contain at least one of 26 medical keywords
6. **Length Filtering**: Documents must be at least 200 words to ensure substantial content
7. **Quality Filtering**:
   - Gopher repetition and quality filters
   - C4 quality filters (excluding terminal punctuation rule)
   - FineWeb custom filters for list-like documents and formatting issues
8. **Token Counting**: GPT-2 tokenizer for token statistics

## Medical Keywords

The dataset employs specialized filtering for medical content using these keywords:

**Core Medical Terms**: medical, diagnosis, treatment, patient, doctor, symptom, therapy, prescription, clinical, healthcare

**Healthcare Facilities**: hospital, clinic, nurse, surgery, pharmacy, pharmaceutical

**Health Conditions**: disease, disorder, condition, medication, drug, vaccine, epidemic, pandemic

**Wellness Terms**: health, wellness

## Data Format

Each example is a JSON object with the following fields:

### Core Fields
- **`text`** *(string)*: The extracted and cleaned text content
- **`id`** *(string)*: Unique identifier from the original WARC record
- **`metadata`** *(dict)*: Extended metadata information

### Metadata Fields
- **`dump`** *(string)*: Common Crawl dump identifier (e.g., "CC-MAIN-2023-50")
- **`dataset`** *(string)*: Dataset identifier ("fineweb-med")
- **`url`** *(string)*: Original webpage URL
- **`date`** *(string)*: Crawl timestamp in ISO format
- **`file_path`** *(string)*: S3 path to source WARC file
- **`language`** *(string)*: Detected language (always "en" for this dataset)
- **`language_score`** *(float)*: Language detection confidence score
- **`token_count`** *(int)*: Number of tokens using GPT-2 tokenizer

## Usage

### Loading the Dataset

```python
from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("pohsjxx/fineweb-med-test")

# Access training split
train_data = dataset['train']

# Example usage
for example in train_data:
    print(f"Text: {example['text'][:100]}...")
    print(f"URL: {example['metadata']['url']}")
    print(f"Tokens: {example['metadata']['token_count']}")
    break
```

### Medical-Specific Filtering

```python
# Filter for clinical documents
clinical_docs = [doc for doc in dataset['train']
                 if 'clinical' in doc['text'].lower()]

# Filter by token count for model training
suitable_docs = [doc for doc in dataset['train']
                 if 512 <= doc['metadata']['token_count'] <= 2048]
```

## Statistics

| Metric | Value |
|--------|-------|
| **Total Documents** | 9 |
| **Total Tokens** | 0 |
| **Average Tokens/Document** | 0.0 |
| **Token Range** | 0 - 0 |
| **Median Tokens/Document** | 0 |
| **Source Dump** | CC-MAIN-2023-40 |
| **Language** | English only |
| **Medical Focus** | Healthcare & medical content |

### Top Content Sources
- **123fish.net**: 4 documents
- **afrigems.de**: 2 documents
- **accg.org**: 1 documents
- **2510000.com**: 1 documents
- **abbybosshair.com**: 1 documents

## Dataset Creation

### Curation Rationale

While FineWeb provides excellent general web text data, specialized domains like healthcare require targeted datasets. FineWeb-Med addresses this need by applying medical-specific filtering to create a high-quality, domain-focused dataset suitable for:

- Training medical language models
- Fine-tuning healthcare AI applications
- Medical text analysis and NLP research
- Healthcare chatbot development

### Source Data

**Primary Source**: Common Crawl web crawl data
- **Dump**: CC-MAIN-2023-40
- **Time Period**: 2023-2024 web crawl
- **Content Type**: Public web pages with medical relevance

### Annotations

We augment samples with automatic annotations:
- **`language`** & **`language_score`**: Generated by FastText language classifier
- **`token_count`**: Calculated using GPT-2 tokenizer

## Considerations for Using the Data

### Social Impact

This dataset enables more accessible development of healthcare AI applications, potentially improving medical text understanding and patient care through better language models.

### Discussion of Biases

The dataset inherits biases from web-sourced medical content, which may reflect:
- Geographic biases in healthcare information availability
- Language biases (English-only content)
- Platform biases from different healthcare websites

### Limitations

- **Code Content**: Limited due to filtering steps; supplement with code-specific datasets if needed
- **Medical Accuracy**: Web content may contain outdated or inaccurate medical information
- **PII Concerns**: Despite anonymization, some personal health information may remain
- **Specialized Domains**: May not cover all medical specialties equally

## Additional Information

### Licensing Information

**License**: Apache 2.0
**Additional Terms**: Subject to Common Crawl's Terms of Use

### Personal and Sensitive Information

We anonymize:
- Email addresses → `email@example.com` or `firstname.lastname@example.org`
- Public IP addresses → Randomly assigned non-responsive IPs

For PII removal requests, please create an issue in our repository.

### Future Work

We plan to expand FineWeb-Med with:
- Additional medical domains and specialties
- Multi-language medical content
- Enhanced quality filtering for medical text
- Integration with medical knowledge bases

## Citation Information

```bibtex
@dataset{fineweb_med,
  title={FineWeb-Med: Medical-Focused Web Dataset},
  author={Generated using datatrove FineWeb methodology with medical filtering},
  year={2024},
  url={https://huggingface.co/datasets/pohsjxx/fineweb-med-test}
}
```

---

*Built with ❤️ using the FineWeb methodology and datatrove*