maximg1/empirical-discourse

Empirical Discourse in Congress

Five sentences from the 116th Congress (2019-20), and a box to type your own. The model calls each sentence empirical or not; for the five, the coders' label sits beside it.

Ilay Abramovich & Maxim German, Workshop on Deep Learning, Tel Aviv University.

What is answering

Three fine-tuned encoders (RoBERTa, DeBERTa, ModernBERT), averaged, with a sentence counted as empirical at 0.60 or above.

That is not the instrument behind the paper's numbers. The published silver labels come from 30 models: 3 encoders x 5 seeds x 2 class-weight arms. This Space runs one seed of the macro-F1 arm, because 30 base transformers is not a demo. The 0.60 operating point was tuned on the full ensemble's out-of-fold predictions, so here it is inherited rather than re-derived, and typed sentences may be scored differently than the reported labels would score them.

The five curated sentences keep their full 30-model scores, which were computed offline on the modern gold set's test half: sentences the ensemble never trained on.

GET /api/meta reports exactly which checkpoints are loaded.

Endpoints

• —GET /api/meta: models loaded, ensemble sizes, threshold and its source.
• —POST /api/score {"text": "..."}: p (ensemble mean), p_std, empirical, per_model.

Source

Weights: maximg1/ced-demo-ensemble, pulled at startup via the DEMO_MODEL_REPO Space variable. The pipeline that produced them is not public yet; the paper is in preparation.