Model Directory

Local inference loads a GLiNER model from a directory on disk. The modelPath you set on a PhEye filter (see Configuration) points at that directory.

Required files

The directory is the layout produced by ph-eye-model-training's ./hub/publish.sh. It contains:

File	Purpose
model.onnx (or onnx/model.onnx)	The exported GLiNER model.
tokenizer.json	The HuggingFace fast tokenizer.
gliner_config.json	Span width, maximum length, and prompt tokens.

model.onnx may sit at the top level of the directory or under an onnx/ subdirectory; both layouts are recognized.

Where to get a model

The published PhEye PII name models on the Hugging Face Hub follow this layout. For example, philterd/ph-eye-pii-en-small, philterd/ph-eye-pii-en-medium, and philterd/ph-eye-pii-en-large each provide the ONNX export, tokenizer, and config. Download a model into a local directory and point modelPath at it.

How the files are used

gliner_config.json drives the inference pipeline: it provides the maximum span width, the maximum sequence length, and the prompt tokens used to build the model input. tokenizer.json is loaded as the fast tokenizer that splits text into subtokens. model.onnx is run by ONNX Runtime. See How It Works for how these come together in the GLiNER pipeline.