laconiq-ai/technique-router-onnx

LaconIQ technique-router (ONNX INT8)

ONNX INT8 export of the technique-router text classifier used by LaconIQ to choose an image-compression technique from a user's query. It runs on CPU with onnxruntime only (no GPU, no PyTorch), in ~32 MB.

Provenance

This is a byte-identical re-host of the upstream ONNX export, with an Apache-2.0 LICENSE, a NOTICE, and this README added so the LaconIQ project has a self-owned, auditable model source.

• Derived from chopratejas/technique-router (Apache-2.0), a fine-tune of microsoft/MiniLM-L12-H384-uncased.
• Original model and ONNX export by Tejas Chopra. Re-hosted unchanged.

Inputs / outputs

• Inputs: input_ids, attention_mask, token_type_ids (int64, max length 64).
• Output: logits over the routing classes; argmax gives the technique, and config.json carries the id2label map.

License

Apache License 2.0. See LICENSE and NOTICE.