Spaces:
Running
Running
| title: DUSUNEN Search Lab | |
| emoji: 🔎 | |
| colorFrom: indigo | |
| colorTo: blue | |
| sdk: static | |
| app_file: index.html | |
| pinned: false | |
| license: mit | |
| short_description: Private, in-browser Turkish semantic search with ONNX | |
| models: | |
| - GoktugD/DUSUNEN-Rota-270M-v1 | |
| datasets: | |
| - GoktugD/DUSUNEN-Retrieval-100K-v1 | |
| tags: | |
| - sentence-transformers | |
| - semantic-search | |
| - turkish | |
| - information-retrieval | |
| # DUSUNEN Search Lab | |
| An interactive, zero-server demonstration of | |
| [`GoktugD/DUSUNEN-Rota-270M-v1`](https://huggingface.co/GoktugD/DUSUNEN-Rota-270M-v1). | |
| The validated mixed-precision model runs locally in the visitor's browser | |
| through ONNX Runtime and WebAssembly. Its embedding table uses int8 while the | |
| transformer blocks remain float; measured mean PyTorch-to-ONNX embedding cosine | |
| agreement is 0.999956. Search text is not sent to an inference API. The app | |
| ranks an original Turkish technology-and-science corpus with cosine similarity. | |
| This demo corpus is separate from both the MS MARCO-derived training set and | |
| the TurHistQuad benchmark. | |
| Measured benchmark results and the complete training recipe are linked from | |
| the model card. Similarity scores are ranking signals, not calibrated | |
| probabilities. | |
| Built with the free Static Spaces runtime; no paid inference endpoint is used. | |