Acknowledge the terms to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Access is granted for non-commercial research and evaluation use only. By requesting access you agree not to redistribute this model or any derivative artifacts, to cite Dhi Technologies in any publication or output that uses it, and to obtain a separate commercial license via dhi-tech.com before any commercial use. Access requests are reviewed manually by Dhi Technologies.

intent2edge-reference-onnx

Reference artifact for the Intent2Edge pipeline, not a production model. This is the ONNX classification export used by the public intent2edge-demo Space. It is a small, fast-loading model file for inspecting what the Intent2Edge pipeline produces from one request: intent, configuration, training, ONNX export and report.

The method and the measured results of the Intent2Edge compiler are described in paper 03, Intent2Edge: Compiling Natural-Language Intent into Refuse-Closed Edge Vision Artifacts (manuscript in preparation). This card does not report an accuracy or latency figure for the toy model in this repository.

What it is

Task: image classification with 3 synthetic classes: red_square, blue_circle, green_triangle.
Backbone: mobilenet_v3_small, ImageNet-pretrained.
Parameters: about 1.52 million. Paper 03 reports 1.52 million parameters for its MobileNetV3-Small classification artifact.
Input: 96x96 RGB images.
Export format: ONNX, opset 17.
Training data: a small synthetic set of colored shapes produced by the pipeline's toy-data command. It is not a natural-image dataset. Matching fixtures are in intent2edge-examples.

Results from paper 03

These are the measured results of the Intent2Edge pipeline. They are not results of this toy artifact.

Prompt-guided versus fixed augmentation. Pooled over five stated deployment conditions, the guided arm scores 5.9 points above the fixed arm on the matched corruption families (bootstrap 95% interval +4.6 to +7.3; 15 dataset-seed pairs). Guided is higher than fixed in all five conditions.
Validity gate. A confidence-calibrated gate at confidence 0.7 lowers false accepts on the seed-23 configurations from 30 to 14, at 1.22 times the refusals.
Export fidelity. The FP32 ONNX export reproduces PyTorch logits to a maximum absolute difference of 2.0e-04, with top-1 agreement of 1.0 on every split.
Certificate retention. On a frozen independent split, 10 of 10 QDQ ONNX graphs keep at least 0.98 of their FP32 accuracy. The certificate holds in 80 of 80 runtime-condition cells on a second host, and in 210 of 210 cells on x86 across 3 ONNX Runtime versions.

What this is not

Not a production or deployment-ready classifier.
Not evidence of real-world classification quality. The labels, images and split are synthetic and small.
Not a benchmark. Use it to inspect the pipeline output and the export format.

Provenance

Code repository: https://github.com/DHI-Technologies-Inc/intent2edge (public, PolyForm Noncommercial 1.0.0)
Commit: ba62c2fd721115f996812af61079ad339a8c5b80
Command as recorded: .venv/bin/python -m prompt2model.cli smoke-test --output-dir output/smoke (the module name in the code repository at that commit)
Date: 2026-07-10