---
license: apache-2.0
base_model: knowledgator/gliclass-v4-edge
library_name: onnx
tags: [onnx, webgpu, game, zero-shot-classification]
---

# Handler: GLiClass for orders

knowledgator/gliclass-v4-edge (built on microsoft/deberta-v3-large) fine-tuned to read the spoken orders in Handler,
a game in which you talk two spies through a mission. These are the ONNX graphs the game runs in the browser with
onnxruntime-web on WebGPU, one run per question:

- `onnx/model_q8f16.onnx` (what the game loads): 8-bit weights (MatMulNBits, weight-only, activations fp16), and the
  word embeddings cut out: it takes `inputs_embeds` (fp16, [batch, length, 1024]) beside `input_ids`, looked up from
  `onnx/embed_tokens_fp16.bin` (128,004 rows of 1024 fp16 values, row = token id).
- `onnx/model_fp16.onnx`: the fp16 export it was made from, taking `input_ids` alone.
- Inputs `input_ids` and `attention_mask` (plus `inputs_embeds` for q8f16); the output is calibrated logits for 25
  option slots. Keep the first n for a question's n options and apply `calibration.json`'s temperature: softmax(logits / T).
- The text is every option as `<<LABEL>>key: description`, then `<<SEP>>`, the question's instructions and the
  state (gliclass 0.1.22's uni-encoder pipeline).

Licence: Apache 2.0, as for GLiClass v4 Edge (Knowledgator); DeBERTa-v3 by Microsoft is MIT.