Skip to main content
This prerelease builds on 0.4.5-beta.3. Install 0.4.5-beta.4 for Cargo and npm, or 0.4.5b4 for Python. Node and WASM remain on npm’s beta tag; latest remains 0.4.4.

PII and general-purpose GLiNER2

Use the existing pinned PII checkpoint or select a compatible general-purpose GLiNER2 span model through a local model manifest. The release checks include fastino/gliner2-privacy-filter-PII-multi and fastino/gliner2-base-v1. The provisioner can generate a versioned manifest and file hashes from explicitly downloaded local assets using --model-id and --revision, or verify an existing manifest using --manifest. Core verifies the selected manifest and listed assets before starting inference. Results report the selected model identity; its revision is included in adapter evidence. Existing legacy PII bundles remain supported with their original compiled asset pins.

Custom entity schemas

GLiNER2 labels are caller-defined, with per-label thresholds and optional descriptions. For example, current_employer produces the typed entity CURRENT_EMPLOYER. The existing phone_number to PHONE mapping is preserved. Empty, padded, duplicate labels and labels mapping to the same output type are rejected. Schemas remain fixed for each session. Custom findings retain byte and codepoint offsets, plus UTF-16 offsets in Node, and work with the existing transformations. scan still routes rule-covered entity types to rules and other selected labels to the model. Use scan_model or Node’s scanModel for raw model output. Downloading and provisioning remain explicit. Inference uses local files through the existing CPU worker and pinned runtime dependencies. This release supports compatible GLiNER2 span models; it does not add classification, relation extraction, GLiNER2.5, GLiNER1, or spaCy adapters. WASM remains rule-based. See the GLiNER2 setup guide and manifest contract.

Release validation

Publication requires DataFog, PII GLiNER2, and general-purpose GLiNER2 to pass packaged Rust and installed Python/Node checks on Linux x64, Linux ARM64, macOS ARM64, and Windows for the release commit. Intel macOS model checks are advisory. These checks validate loading, response contracts, routing, offsets and lifecycle behavior; they do not make model-accuracy claims.