GLiNER2.5 small Core ML classification

This is a Core ML conversion of the classification decision path of Fastino's GLiNER2.5 small, source revision 7e6f537f10337497069276892a5ef435028252ce. The original model has 73,881,879 parameters and is licensed Apache-2.0. This export contains its encoder and trained classification head (70,944,385 parameters); entity, relation, record, count and span extraction heads are not included. Use the original checkpoint for those tasks.

The FP16 L128/K8 package occupies 151,542,752 bytes. It accepts up to eight classification labels, with native GLiNER2 schema rendering. The tokenizer files are included in this repository. The Core ML deployment target is iOS 17 or macOS 14. preprocessing.py and runtime.py implement the request path without loading the original model weights at inference time.

A Python example, from the directory containing this README:

uv sync
uv run python runtime.py --model-dir . \
  --text "The rocket launched successfully." \
  --task topic --labels '["science","sports","politics"]'

The model returns a label, confidence and probabilities. Inputs that exceed the bucket capacity need a larger bucket; the runtime must not truncate them. The runtime requires macOS to execute Core ML prediction.

Validation

On an Apple M5 Pro running macOS 27.0, the FP16 package matched the native classifier's chosen label on all 100 eligible requests selected in source order from a fixed application suite. The largest chosen-label confidence difference was 0.001723. Median Python MLModel.predict call time was 7.72 ms, including Python/Core ML dispatch; this is not an ANE-only measure. There were 300 rows with more than eight options and seven over-length rows before 100 eligible rows were collected. The smoke run does not establish a Decision Index score or broader application accuracy. verify-application100.json contains the exact counts and results.

An experimental LUT8 per-tensor package occupies 76,211,068 bytes and matched all 100 chosen labels, but its largest confidence difference was 0.07859. It is not the recommended parity artifact. Grouped-channel LUT8 requires an iOS 18 or later Core ML target and has not yet been evaluated.

An experimental token-embedding-only int8 package occupies 102,771,476 bytes, 32.2% less than FP16. It matched the native classifier's chosen label on the same 100 eligible application requests (largest chosen-label confidence difference 0.01685). A paired 14-request FP16 comparison kept every choice with worst absolute probability difference 0.01758 on CPU+ANE. Full-request CPU+ANE medians were 3.701 ms FP16 and 3.719 ms int8 on battery power; one small AC-powered paired check measured 3.702 and 3.668 ms. The results are effectively tied, so this is a size option, not a speed claim. The package retains L128/K8 classification scope and is selected explicitly:

uv run python runtime.py --model-dir . \
  --package gliner2_small_classification_embedding_w8_L128_K8.mlpackage \
  --text "The rocket launched successfully." \
  --task topic --labels '["science","sports","politics"]'

The package file hashes and selected-request reports are under reports/.

Conversion source, pinned dependencies and verification scripts are included. The exporter freezes two static DeBERTa attention expressions and uses a finite FP16 mask sentinel; the real-model tests and native/Core ML checks guard this rewrite. Upstream authors receive credit for the original model; Fluid Inference performed this Core ML conversion.