foxmind / README.md

Name the model foxmind; keep the GLiNER2 base model credit

ba96433 verified 9 days ago

1.71 kB

license: apache-2.0
base_model: fastino/gliner2-multi-v1
library_name: transformers.js
tags:
  - gliner2
  - foxpilot
  - onnx
  - token-classification

foxmind

foxmind is the on-device decision model of foxpilot, a browser agent for Firefox. It is an ONNX export of the agent graph of GLiNER2 multi-v1, with a batch axis.

Base model: fastino/gliner2-multi-v1 (Apache-2.0), by Fastino.

It is the same single-graph export as onnx-community/gliner2-multi-v1-agent-ONNX, with one change: every input and output has a batch axis.

Name	Shape
`input_ids`, `attention_mask`	`[batch, seq]`
`word_positions`	`[batch, words]`
`schema_positions`	`[batch, schema]`
`cls_logits`	`[batch, schema - 1]`
`count_logits`	`[batch, 20]`
`span_logits`	`[batch, schema - 1, words, 8]`

At batch 1 it behaves like the original export. Pad shorter rows with 0 and set their mask to 0.

Files

onnx/model.onnx + model.onnx_data: fp32
onnx/model_fp16.onnx + model_fp16.onnx_data: fp16

Accuracy

export/verify_onnx.py in the foxpilot repository checks this export. In fp32, 14 reference calls match the Python library, and each padded batch row matches its single call within 8.0e-6 on label and span scores.

Speed

Firefox 157, WebGPU, fp32: one call with 4 prompts takes 517 ms, against 1,242 ms for 4 single calls (2.4x).

License

Apache-2.0, the license of the base model. Export scripts: pooriaarab/foxpilot (MIT).