Eyes & Ears: speaker attribution and character finder (Core ML)

The models behind the Ensemble cast in the Eyes & Ears iOS app, a free, non-commercial read-along ebook reader. They decide who speaks each line of dialogue and find a book's characters, on the device.

  • DeezerEncoder.mlmodelc: ModernBERT-large encoder, retrained with Deezer's ModernBookNLP recipe, fp16, enumerated input lengths (compiled Core ML model).
  • deezer-head.bin: the pair-scoring head (batch norm folded in), float32 little-endian.
  • finder-heads.bin: the character finder's two small heads (name spotter, same-person classifier) and its threshold, float32 little-endian. They only work on this encoder.
  • tokenizer/: ModernBERT's tokenizer files.

How it was trained

  • Encoder and attribution head: Deezer's training recipe on 16 PDNC novels. The Sign of the Four, Pride and Prejudice and The Mysterious Affair at Styles were held out as a test set. These are not Deezer's released weights.
  • Name spotter: LitBank, PDNC, and public-domain Project Gutenberg novels labelled with Claude for this project.
  • Same-person classifier: PDNC and those labelled novels.

License and credits

These weights are offered under CC BY-NC 4.0, for non-commercial use.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for crozierm/eyes-and-ears-attribution

Finetuned
(437)
this model