GLiNER2.5 base (Re-DocRED spans) — ONNX for WebGPU
ONNX export of raphaelmansuy/gliner2.5-base-redocred-spans for onnxruntime-web / WebGPU in the edgextract browser demo.
Fine-tuned on Re-DocRED train mention spans (PER, ORG, LOC, TIME, NUM, MISC). Threshold used in the kept product row: 0.40.
Files
| Path | Role |
|---|---|
onnx/model.onnx |
Encoder + boundary proposer + pair reranker (~741 MB) |
export_config.json |
Export metadata (opset 17, candidate budget 192) |
| tokenizer files | Same vocab as the PyTorch checkpoint |
Browser
The edgextract Wikidata mode loads this graph with WebGPU (WASM fallback), packs entity labels in JS (Pastel GlinerBoundaryRuntime), and keeps spans with sigmoid(pair_logit) >= 0.40. Relation scores still come from ATLOP ONNX (raphaelmansuy/atlop-roberta-base-onnx).
Export tooling adapted from Pastel-Org/gliner2.5-onnx-webgpu (MIT). Base architecture: Fastino GLiNER2.5 (Apache-2.0).