| --- |
| title: Data-use Extraction — WebGPU |
| emoji: 🐈 |
| colorFrom: indigo |
| colorTo: blue |
| sdk: static |
| sdk_version: 1.0.0 |
| pinned: false |
| license: apache-2.0 |
| short_description: In-browser data-use extraction on WebGPU |
| --- |
| |
| # Data-use Extraction — WebGPU |
|
|
| Fully client-side demo that extracts data mentions from research text and |
| attributes each one with provenance and usage/impact — all running locally in |
| your browser on WebGPU. No server, no uploads; nothing leaves your machine. |
|
|
| ## Pipeline |
|
|
| ``` |
| text |
| → span model (GLiNER) → mention spans + specificity |
| → LFM2.5-350M multitask → provenance {producer, year, geography, acronym} |
| + usage/impact {data_type, usage_action, impact_label, usage_summary} |
| → verbatim filter → render |
| ``` |
|
|
| ## Models |
|
|
| | stage | model | HF repo | |
| |---|---|---| |
| | span detection | GLiNER | [`rafmacalaba/gliner_datause_extended-onnx`](https://huggingface.co/rafmacalaba/gliner_datause_extended-onnx) | |
| | attribution | LFM2.5-350M multitask | [`rafmacalaba/lfm2.5-350M-datause-multitask-onnx`](https://huggingface.co/rafmacalaba/lfm2.5-350M-datause-multitask-onnx) | |
|
|
| Models download from the Hub on first load and are cached in your browser's |
| IndexedDB. Requires a WebGPU-capable browser (Chrome/Edge with WebGPU enabled). |
|
|
| ## Run locally |
|
|
| ```bash |
| git clone https://github.com/rafmacalaba/ai4data-playground.git |
| cd ai4data-playground/web && python3 -m http.server 8000 |
| # open http://localhost:8000 |
| ``` |
|
|
