| --- |
| title: README |
| emoji: 🧬 |
| colorFrom: indigo |
| colorTo: blue |
| sdk: static |
| pinned: false |
| --- |
| |
| # ExeQut |
|
|
| We build data and AI systems for research and government, with a focus on biomedical data |
| infrastructure, search, and knowledge graphs. |
|
|
| Our work covers the full path from raw source data to something a researcher can actually find: |
| ingestion and transformation pipelines, entity resolution, knowledge graphs, search platforms, and |
| the domain models that make specialised data discoverable. |
|
|
| ## What we publish here |
|
|
| Models we build in house for biomedical text and metadata. We care about two things in particular, |
| and both show up in how we evaluate: |
|
|
| - **Honest measurement.** We score against human-labelled ground truth, hold the evaluation set out |
| of training, and report single-threshold results rather than tuned upper bounds. |
| - **Current vocabulary.** Biomedical language moves. A model trained on a frozen vocabulary quietly |
| stops recognising the terms researchers are actually using. |
|
|
| ## Models |
|
|
| - **[ExeQut MeSH Tagger](https://huggingface.co/Exequt/mesh-tagger)** - biomedical MeSH tagging |
| across 25,489 current MeSH descriptors, trained from scratch on roughly 3 million indexed MEDLINE |
| records. It reaches 0.663 micro-F1 against expert human indexing, ahead of the widely used open |
| model for this task. |
|
|
| ## Work with us |
|
|
| We take on data platform, search, and applied machine learning work in biomedical and public sector |
| settings. Get in touch at [exequt.com](https://exequt.com). |
|
|
