Classify and extract text with GLiNER2.5 Decide

This guide is for developers who need local classification and schema driven extraction. The managed English checkpoint runs its DeBERTa encoder and classification, count, and span heads in Swift/MLX.

Pull the pinned checkpoint:

mere.run model pull text-classify-gliner25-decide

Classify text

Save a request as classification.json:

{
  "text": "The package arrived late, and I need a refund.",
  "tasks": [
    {"name": "intent", "labels": ["order_status", "refund_request", "cancel_order"]},
    {"name": "topics", "labels": ["shipping", "billing", "account"],
     "multi_label": true, "threshold": 0.4}
  ]
}

Check the token count, then classify the text:

mere.run text classify --input classification.json --preflight --pretty
mere.run text classify --input classification.json --pretty

Each task returns selected labels and a probabilities map. Single label tasks use softmax; multi label tasks use sigmoid and the task's threshold. If no multi label score reaches the threshold, the highest scoring label is returned. Optional prompt and descriptions fields add task instructions and label definitions.

Extract entities, relations, and structures

Save a request as extraction.json:

{
  "text": "Alice Smith joined Acme in Paris in 2024.",
  "entities": [{"name": "person"}, {"name": "organization"}, {"name": "location"}],
  "relations": [{"name": "works_for"}, {"name": "located_in"}],
  "structures": [{
    "name": "employment",
    "fields": [
      {"name": "person", "multiple": false},
      {"name": "organization", "multiple": false},
      {"name": "location", "multiple": false}
    ]
  }],
  "threshold": 0.5
}

Run extraction:

mere.run text extract --input extraction.json --preflight --pretty
mere.run text extract --input extraction.json --pretty

The result groups entity spans, relation pairs, and structure records. Each span includes text, confidence, and character offsets. Entity and relation terms can include a description. Structure fields can include a description and multiple (true by default) to limit a field to one span or keep several. Add classifications with the same task objects used by text classify to run a joint schema in one encoder pass.

Process batches and long documents

Pass --batch to read a JSON array of independent requests and return an array in the same order. Pass --long to split each document into overlapping word chunks and merge its results. Both flags work with text classify and text extract:

mere.run text classify --input classifications.json --batch --long
mere.run text extract --input extractions.json --batch --long

Long extraction returns offsets into the original document. Overlapping chunks can report the same span; the runtime keeps one copy. The default chunk size is 384 words with 64 words of overlap. Each encoded chunk still must fit the checkpoint's 512 token limit, including the schema. The runtime shortens a chunk when the schema needs more tokens.

The loopback API accepts the same single request at POST /v1/text/classifications or POST /v1/text/extractions with a model field. Set long: true for long documents. For a batch, send a requests array with model and optional long; the API returns an array.

In macOS Studio, open Text > Classify for editable label tasks or Text > Extract for editable entities, relations, structures, and joint classification tasks. Both pages support Check fit, structured results, and JSON import and export. Turn on Process long text in overlapping chunks for a long document.

Check the limits

The checkpoint specializes in English classification. Its retained count and span heads produce entity, relation, and structure predictions, but the published model card does not qualify their accuracy. The native checkpoint tests compare classification scores and representative extraction predictions with the Python reference. Evaluate extraction quality on your own data.

The normal request must fit within 512 subword tokens. Classification accepts up to 16 tasks and 64 labels per task. Scores depend on the supplied schema and are not calibrated decision probabilities. The model does not generate text or explanations. The separate 1B and multilingual Decide checkpoints are outside this managed model entry.

The upstream Python library also offers schema metadata such as validators, choice fields, span attributes, and record formation policies. This native request implements the checkpoint's classification, entity, relation, and structure heads with basic field controls; it does not expose those advanced Python metadata policies.

Sources and license

• Pinned GLiNER2.5 Decide model snapshot
• GLiNER2 source
• DeBERTa implementation

The model card declares Apache 2.0. The managed catalog pins the model snapshot revision.