docs/gliner25-decide.md
GLiNER2.5-Decide
GLiNER2.5-Decide is a System-1 decision classifier. A DeBERTa-v3-large
encoder (already implemented for GLiNER2.5) extracts label embeddings, and
a classification head (Linear(H, 2H) → ReLU → Linear(2H, 1)) scores each
label in a single forward pass. It answers choice, score, and noul
question types through /v1/systemone.
GLiNER2.5-Decide runs through the shared paths, so the usage guide
covers the /v1/systemone request shape. This page carries the checkpoint
and what has not been measured.
The checkpoint#
| field | value |
|---|---|
| repo | fastino/GLiNER2.5-Decide |
| format | safetensors, F32, 486M parameters |
| architecture | SpanExtractor (same name as MODEL-GLINER25, with a classification head) |
The encoder config is read from encoder_config/config.json (DeBERTa v2).
The classifier weights load from the same safetensors checkpoint.
Run it#
build/examples/vllm-server --model /path/to/GLiNER2.5-Decide --port 8000
Then send a systemone request:
curl http://localhost:8000/v1/systemone \
-H 'Content-Type: application/json' \
-d '{"state":"john works at google",
"questions":{"pick":{"type":"choice","instructions":"entity type",
"criteria":{"person":null,"organization":null}}}}'
The sequence layout is [CLS] text [SEP] [P] task [L] label1 [L] label2 ... [SEP]. The encoder runs once; label embeddings are extracted at [L]
positions and scored by the classifier. Confidence uses the shared kev
formulas (ChoiceConfidence, ScoreConfidence).
What has not been measured#
No token gate result exists for this checkpoint yet. No byte has been fetched or hashed. The golden tests (classification head, softmax/sigmoid) run on synthetic weights.
GPU (CUDA) build verification, GGUF k-quants, E2E parity vs the GLiNER2
library oracle, and the constraint-feasibility decoding pipeline (exact
and beam decoders) are owed.