lelloman/bert-torrent-classifier

BERT Torrent Classifier

A fine-tuned BERT-tiny model for classifying torrent content into media types.

Model Details

• Base model: prajjwal1/bert-tiny
• Task: Multi-class text classification
• Labels: audio, video, software, book, other
• Format: ONNX (with embedded weights)
• Size: ~17MB

Training

• Training data: ~10k torrent names with 4-LLM consensus voting
• LLM ensemble: qwen2.5:3b, gemma3:4b, mistral:7b, qwen3-coder:30b
• Consensus rules: 4-agree = high confidence, 3v1 = majority vote, 2v2 = discarded
• Accuracy: ~92% on held-out test set

Usage

This model is designed for use with mimmo, a Rust library for torrent content classification. The ONNX model is embedded directly in the binary at compile time.

// Model is automatically downloaded during build
const MODEL_BYTES: &[u8] = include_bytes!("../models/bert/model_embedded.onnx");
const TOKENIZER_JSON: &str = include_str!("../models/bert/tokenizer.json");

Performance

• Inference: <10ms per sample (CPU)
• Used as ML fallback when pattern matching is inconclusive

Files

• model_embedded.onnx - ONNX model with embedded weights
• tokenizer.json - HuggingFace tokenizer
• vocab.txt - Vocabulary file
• config.json - Model configuration

License

MIT