GLiNER2.5-Decide-CoreAI
Core AI (.aimodel) conversion of the classification path of Fastino's GLiNER2.5-Decide.

Origin
  Model:     fastino/GLiNER2.5-Decide
  Source:    https://huggingface.co/fastino/GLiNER2.5-Decide
  Revision:  7ee5da4c2415e32259bcdc0b1a7367c32ce8d6f6 (2026-09-24)
  Weights:   model.safetensors, sha256 40a5a23ff860dc3dff426cecd1048cacdd29c648c96db209dad818e9686dc997
  Provider:  Fastino
  License:   Apache License 2.0 (declared on the model card; the source repository has no LICENSE file)
  Base:      microsoft/deberta-v3-large (MIT License)

Software used
  gliner2 2.0.0 (Apache License 2.0): the fp32 reference the conversion was checked against.
  transformers 4.57.6 (Apache License 2.0): DebertaV2Model, into which the encoder weights were loaded.

Conversion
  The released fp32 weights were loaded strictly into transformers' DebertaV2Model and the classification head
  (Linear 1024->2048, ReLU, Linear 2048->1) and exported as one static graph with Apple coreai-torch 0.4.1 /
  coreai-core 1.0.0b2, at two sequence lengths (256 and 512). The bundles store the released weights in half
  precision. No weights were retrained, pruned or quantized. The span and count heads, which classification does
  not use, are not in the graph. DeBERTa-v3's relative-position bucket table is computed in PyTorch at export time
  and stored in each graph as a constant. The ios/ bundles are the same graphs compiled ahead of time with
  coreai-build for the iPhone 18 Pro (h19p). Converted and published by mlboydaisuke
  (https://huggingface.co/mlboydaisuke), 2026-09.

Changed files
  tokenizer/tokenizer_config.json is the source file with one change: tokenizer_class DebertaV2Tokenizer ->
  XLMRobertaTokenizer, so that swift-transformers loads the same SentencePiece Unigram model. tokenizer.json and
  special_tokens_map.json are unchanged. source/config.json and source/encoder_config.json are the source
  repository's config.json and encoder_config/config.json, unchanged.

Test fixtures (gate/)
  readme21.json: the 21 classify_text examples of the model card at the revision above.
  fast_decisions_s256.json, fast_decisions_long.json: rows of fastino/fast-decisions (Apache License 2.0), revision
  1a33070cabf94ce2e29105482dd2ef6c157ad7f2.
  Each case carries gliner2 2.0.0's fp32 inputs, logits and decisions.

This distribution is provided under the Apache License 2.0 (see LICENSE). When redistributing any part of it,
keep this NOTICE and the LICENSE file with it.