| language: en | |
| tags: | |
| - torrent | |
| - usenet | |
| - metadata-extraction | |
| - title-extraction | |
| - span-prediction | |
| - onnx | |
| license: mit | |
| # MPE Engine v4.0 (Metadata Probability Extraction Engine) | |
| This model is a token classification and span extraction transformer designed to parse release names from Torrents and Usenet. | |
| It uses an **Atomic Normalization Layer** to compress relational metadata (like `Season 1-5 Complete` or `S01E02-E04`) into single, unbreakable tokens before passing them to a MiniLM encoder. | |
| ## Architecture | |
| - **Backbone:** sentence-transformers/all-MiniLM-L6-v2 | |
| - **Head 1 (Binary):** Is the token metadata? (BCE Loss) | |
| - **Head 2 (Category):** What type of metadata? (Cross-Entropy Loss) | |
| - **Head 3 & 4 (Span):** Start/End position of the Title (Cross-Entropy Loss) | |
| ## Capabilities | |
| - Extracts titles cleanly from Western, Anime, K-Drama, and Indian OTT release styles. | |
| - Resists breaking on ambiguous titles like `1923`, `It`, `DV`, `From`. | |
| - Supports exhaustive ISO languages (3-letter codes like `tam`, `kor`). | |
| - Handles complete pack ranges, daily episodes, and multi-episode ranges. | |
| ## Files Included | |
| - `mpe_model.onnx`: The production-ready ONNX export for <10ms CPU inference. | |
| - `pytorch_model.bin`: The native PyTorch weights. | |
| - `tokenizer/`: The WordPiece tokenizer vocabulary. | |
| - `categories.json`: The hierarchical category mapping. | |