Skip to content

GLiNER2.5-Decide

  • Adapter: gliner2-decide — fused graph with [L] label markers and [SEP_STRUCT]
  • License: Apache-2.0
  • Parameters: 340M · Architecture: DeBERTa-v3-large · Language: English only

Default variant
Repo onnx-community/GLiNER2.5-Decide-ONNX
Files onnx/model_q4f16.onnx, onnx/model_q4f16.onnx_data, tokenizer.json
Quantization q4f16 (4-bit MatMul, fp16 elsewhere)
Download 523.5 MB
Role default

Variants

Variant Quantization Size Role Install
q4f16 * q4/fp16 523.5 MB default dtk models download gliner2.5-decide --variant q4f16

This is the only catalogued and verified build. There is no fp16 or fp32 variant in the toolkit; see Variants and quantization.

Reproducibility

  • Revision: 18271032588167fe9a3fbb8c5a4900b1231f8119
  • Default file hashes (SHA256):
  • onnx/model_q4f16.onnx — 93f3e0d4ed8bbd4858840f8a2e66028a70e86a482e1ff9822d2ac726381684f6
  • onnx/model_q4f16.onnx_data — df2b05dc97ff875e0d2b51ec21de0049b7b3e7feddaa5224d318dc79e66210e8
  • tokenizer.json — 3ad87d9ffe669147063e70850927dd2da90249e2acc5c8527f1eb65df467bcc8
  • Verify: dtk models verify gliner2.5-decide --variant q4f16
  • Bench: uv run python bench/run.py --suite bench/suites/smoke.jsonl --models gliner2.5-decide

Evidence (2026-10-19)

  • Model-card refund example → refund_request 0.944 (the card expects refund_request).
  • Flight-change example → seat_change 0.593; change second at 0.334 (the card's potential output was change).
  • Sentiment example → positive 0.99; the card's potential output was mixed. Documented as a known divergence, not hidden.
  • 912 ms/decision on CPU.

Contract notes

  • One fused graph: input_ids, attention_mask, marker_positions → logits per label.
  • Sequence: ( [P] prompt [DESCRIPTION] … ( [L] label … ) ) [SEP_STRUCT] [SEP_TEXT] words.
  • The state is lowercased and tokenized word by word; labels keep their case.
  • noul uses labels yes / no; score uses the level strings as labels.

Known limitations

  • English only (declared by the export).
  • Quantized q4f16; the fp16/fp32 builds may resolve the sentiment example differently.
  • The model card's examples are "potential outputs", not guarantees; validate on your data.

Loading each variant

REST selects the quantization through the standard model field using the alias gliner2.5-decide@variant; there is no variant field in the request body.

q4f16 (q4/fp16, default)

Library

from decision_toolkit import DecisionModel

model = DecisionModel("gliner2.5-decide", variant="q4f16")

CLI

dtk models download gliner2.5-decide --variant q4f16
dtk models use gliner2.5-decide --variant q4f16
dtk decide --model gliner2.5-decide --variant q4f16 \
  --state "..." --q 'urgent=noul:Is this urgent?'

REST

curl -s http://127.0.0.1:8000/v1/systemone \
  -H "Content-Type: application/json" \
  -d '{"model": "gliner2.5-decide@q4f16", "state": "...",
       "questions": {"urgent": {"type": "noul",
                     "instructions": "Is this urgent?"}}}}'

Usage

dtk models download gliner2.5-decide
dtk decide --model gliner2.5-decide --state "My subscription renewed after the outage…" \
    --q intent='choice:refund_request|order_status|cancel_subscription|other'

Provenance

ONNX Community export of fastino/GLiNER2.5-Decide; classification path only (extraction heads are not exported).