lelloman/bert-torrent-nsfw

BERT Torrent NSFW Classifier

A BERT-tiny binary classifier for detecting NSFW content in torrent names and file listings.

Model Details

• —Architecture: BERT-tiny (2 layers, 128 hidden, 2 attention heads)
• —Task: Binary classification (safe vs nsfw)
• —Input: Torrent name + top 3 filenames by size
• —Output: Logits for [safe, nsfw]
• —Format: ONNX (embedded weights)
• —Size: ~17MB

Usage

This model is designed to be used with mimmo, a Rust-based torrent content classifier.

# With mimmo CLI
mimmo --detect-nsfw "Torrent.Name.Here"

Training

Trained on a dataset of ~10k torrent names labeled as safe/nsfw using:

• —BERT-tiny base model
• —Binary cross-entropy loss
• —AdamW optimizer

Performance

The model is used as a fallback after keyword and pattern matching stages:

1. 1.Keywords - HashSet lookup for explicit terms (confidence: 0.95)
2. 2.Patterns - Regex for JAV codes, studio patterns (confidence: 0.90)
3. 3.ML (this model) - For ambiguous cases

Limitations

• —Trained primarily on English torrent names
• —May not catch all NSFW content, especially with obfuscated names
• —Should be used in combination with keyword/pattern stages for best results

License

MIT