Module gliner 

GLiNER — zero-shot span NER on the engine’s DeBERTa backbone.

The backbone (disentangled attention) is the engine’s; everything here is the head that turns its per-token states into typed entity spans. The path, taken from the real checkpoint (urchade/gliner_small-v2.1) and gated against the gliner package layer by layer:

input:  <<ENT>> type₁ <<ENT>> type₂ … <<SEP>> word₁ word₂ …

DeBERTa hidden states                                   [T, 768]
  → projection                       Linear(768 → 512)  [T, 512]
  → gather   prompts at the <<ENT>> positions           [C, 512]
             words at each word's FIRST subtoken        [W, 512]   (subtoken_pooling = "first")
  → BiLSTM   1 layer, 256/dir, concat                   [W, 512]
  → SpanMarkerV0:  start = project_start(words)
                   end   = project_end(words)
                   span[L][k] = out_project(relu(start[L] ‖ end[L+k]))
                                                        [W, max_width, 512]
  → prompt_rep(prompts)                                 [C, 512]
  → scores[L][k][c] = span[L][k] · prompt[c]            [W, max_width, C]
  → sigmoid, threshold, flat greedy overlap resolution  → entities

Every MLP is GLiNER’s create_projection_layer: Linear(d → 4·out) → ReLU → Dropout → Linear. Dropout is inference-time identity, which is why only the .0. and .3. weights exist.

Span (L, k) means the word range [L, L+k] inclusive; it is scored only when L+k < W.

Structs

Entity

One predicted entity: a WORD range [start, end] (inclusive) with its type and probability.

Gliner

GLiNER: the engine’s DeBERTa backbone plus the span head.

TextEntity

One predicted entity over the ORIGINAL text: byte offsets, the surface string, and its type.

Enums

GlinerDevice

Which device GLiNER runs on — backbone AND head together (they share one adapter, so the hidden states never round-trip between devices).