SmolLM Torrent Metadata Extractor

Fine-tuned SmolLM-360M-Instruct for extracting structured metadata from torrent names.

Model Description

This model extracts title, artist, and year information from torrent filenames for audio and video content.

Input Format

<|im_start|>user
<|extract|>[content_type] torrent_name<|im_end|>
<|im_start|>assistant

Content Types

  • audio/album โ†’ outputs: album | artist | year
  • audio/track โ†’ outputs: track | artist | year
  • video/movie โ†’ outputs: title | year
  • video/episode โ†’ outputs: series_title
  • video/season โ†’ outputs: series_title
  • video/series โ†’ outputs: series_title

Examples

Input: <|extract|>[audio/album] Pink Floyd - The Dark Side of the Moon (1973) [FLAC] Output: The Dark Side of the Moon | Pink Floyd | 1973

Input: <|extract|>[video/movie] The.Matrix.1999.1080p.BluRay.x264-GROUP Output: The Matrix | 1999

Training

  • Base model: SmolLM-360M-Instruct
  • Method: LoRA fine-tuning (r=16, alpha=32)
  • Training data: ~200k samples from Spotify catalog validation
  • Checkpoint: 11500 steps

Files

File Size Description
smollm-f32.gguf 1.4GB Full precision GGUF
smollm-q4_k_m.gguf 259MB Q4_K_M quantized (recommended for inference)

Usage with llama.cpp

./llama-cli -m smollm-q4_k_m.gguf -p "<|im_start|>user
<|extract|>[audio/album] Radiohead - OK Computer (1997) [FLAC]<|im_end|}
<|im_start|>assistant
" -n 32

License

MIT

Downloads last month
8
GGUF
Model size
0.4B params
Architecture
llama
Hardware compatibility
Log In to add your hardware

4-bit

32-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for lelloman/smollm-torrent-metadata

Quantized
(23)
this model