Access is granted for non-commercial research and evaluation use only. By requesting access you agree not to redistribute this dataset or any derivative data, 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 Toy Examples

Product: Intent2Edge, a compiler that turns a natural-language edge-vision request into a trained ONNX candidate with a recomputable proof bundle. The pipeline runs a typed sequence: prompt, dataset config, training, calibration with conformal abstention, ONNX export, an optional compression step with an accuracy-floor gate, and a hard-case capture store.

What this is (and isn't)

This is not a benchmark dataset. It holds the small synthetic fixtures that exercise the pipeline end to end, generated by the code repository's toy-data command, plus one pipeline output file for those fixtures. The method is evaluated in paper 03 on five public image-classification datasets, and the measured results are summarized below.

The fixtures let a reader check the pipeline's inputs and its output format against the actual files.

What's in this dataset

classification/{red_square,green_triangle,blue_circle}/: 12 procedurally drawn 128x128 PNGs per class (36 images).
detection/images/: 12 PNGs, image_001.png through image_012.png, 128x128 each, plus detection/annotations.json, a COCO-style toy detection set with 2 categories (square, circle) and 24 bounding boxes (2 per image).
smoke_test_results.json: the pipeline's output file for these fixtures (training, calibration, ONNX export and runtime check). It is included as a format example. Its values come from a toy-scale run and are not reported as results anywhere in this repository.

How to load it

import json
from huggingface_hub import hf_hub_download

repo_id = "Dhi-Technologies/intent2edge-examples"
ann_path = hf_hub_download(repo_id, "detection/annotations.json", repo_type="dataset")
smoke_path = hf_hub_download(repo_id, "smoke_test_results.json", repo_type="dataset")

annotations = json.load(open(ann_path))
smoke = json.load(open(smoke_path))

print(len(annotations["images"]), "images,", len(annotations["annotations"]), "boxes")
print(smoke["classification"]["metrics"]["calibration"].keys())

To fetch a classification image, use the same call, for example hf_hub_download(repo_id, "classification/red_square/red_square_000.png", repo_type="dataset").

Results

The measured results of Intent2Edge are in paper 03, Intent2Edge: Compiling Natural-Language Intent into Refuse-Closed Edge Vision Artifacts (manuscript in preparation). Headline results:

Prompt-guided versus fixed augmentation. Pooled over five stated deployment conditions, the guided arm scores 5.9 points above the fixed arm on matched corruption families (bootstrap 95% interval +4.6 to +7.3; 15 dataset-seed pairs). Guided is higher in all five conditions.
Validity gate. On 742 trained configurations, 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 and retention. The FP32 ONNX export reproduces PyTorch logits to a maximum absolute difference of 2.0e-04. On a frozen independent split, 10 of 10 QDQ ONNX graphs keep at least 0.98 of their FP32 accuracy.

The fixtures in this repository are not used for those results.

Reproducing the pipeline run on these fixtures

PYTHONPATH=src python -m prompt2model.cli generate-toy-data --task all --output-dir output/toy_data
PYTHONPATH=src python -m prompt2model.cli smoke-test --output-dir output/smoke

The module name is as it appears in the code repository at the commit used for this release.

Pipeline behavior that the fixtures exercise

The pipeline seeds torch and random from config.dataset.seed (default 42).
The train, validation and test split is stratified by class, so every class is present in each split.
The compression step keeps the uncompressed model unless the compressed model holds the accuracy floor (relative floor 0.98 by default).

Limitations

Toy scale only: 36 classification images and 12 detection images. These fixtures check that the pipeline runs end to end. They do not measure model quality.
The fixtures contain no held-out real-world evaluation set.

License

This dataset (fixtures and pipeline output) is released under CC BY-NC 4.0 (non-commercial). Access is gated and requires manual approval. It is provided for non-commercial research and evaluation only. Redistribution is not permitted, and any publication or output using it should cite Dhi Technologies. Commercial use requires a separate agreement; contact dhi-tech.com. The code repository linked above is licensed separately under the PolyForm Noncommercial License 1.0.0 (commercial use: https://dhi-tech.com/pricing/).