What problem does it solve? Tagging regions and spans of text (document sections, structural markers, field values) usually means brittle regex or expensive prompt-based extraction. This Skill guides fine-tuning the OpenAI Privacy Filter (OPF), a small bidirectional token-classifier, to detect arbitrary custom span categories with a trained model instead. ## Core Features & Use Cases - Custom label-space fine-tuning: Define your own span ontology (JSON label space + JSONL annotations with character offsets) and train OPF via its opf train CLI, with replace-vs-union guidance for keeping native PII categories. - Ontology design patterns: Anchor-based schemes (single-anchor and start/end pairs) that work around OPF's banded attention so long regions are reconstructed in post-processing rather than dense-labeled. - Annotation tooling and validation: A bundled opf_annotate.py script validates offsets, overlaps, and label-space coverage, converts match-based spans to offsets, and previews boundaries inline. - Use Case: Segment Brazilian legal decisions into relatório, fundamentação, and dispositivo sections by fine-tuning OPF on a few thousand PT-BR annotated examples, then extracting the operative outcome for downstream statistics. ## Quick Start Ask the agent to design a span label space and fine-tune OPF on your annotated JSONL training file to tag custom regions in your documents.
