seanll95/excerpt-highlight-scorer

Excerpt highlight scorer

Core ML cross-encoder that ranks candidate passages by how likely a reader is to highlight them. Used by the Excerpt iOS app's Suggested Highlights feature, which downloads it on first use rather than shipping it in the app.

Layout

Each revision is a top-level directory. The app pins one by name (SuggestedHighlightModels.modelRevision) and downloads only that directory, so a new upload can never be mistaken for a copy already on disk.

v1/electra-crowd-v4.mlpackage/   # Core ML package, compiled on device
v1/tokenizer.json                # WordPiece vocabulary
v1/tokenizer_config.json         # provenance only
v1/manifest.json                 # provenance only

Publishing a replacement means uploading v2/ and bumping the constant in the app. Existing installs keep working against the revision they were built for.

Inputs and outputs

input_ids, attention_mask, token_type_ids → logits. Batch dimension is fixed at 1; the app batches through MLArrayBatchProvider.