Shared Task

Why representations

Work across the NLP pipeline profits from robust style representations: curating stylistically diverse training and evaluation data, steering style in generation, and applications like machine-text detection. The shared task therefore targets representations directly.

It uses STEB (Rivera Soto et al., 2026), which standardizes a fragmented evaluation landscape across 96 datasets and 7 languages. Beyond the released benchmark, we also evaluate submitted models on private datasets covering new domains and applications, to test how well representations generalize.

Tracks

Participants may enter either of two complementary tracks.

(a) Application-oriented style representations

Target one or more applications, such as authorship retrieval or machine-text detection, or aim for broad performance across stylistic tasks. Task-specific systems are ranked by the macro-average over the datasets of the selected application. General-purpose systems are ranked by STEB’s operational score, which aggregates performance across clusters of correlated datasets.

(b) Linguistically grounded style representations

Following the definition of style in Wegmann et al. (2026), systems are ranked by STEB’s definitional score. It evaluates whether embeddings encode relevant linguistic features, capture stylistic patterns across genres, registers, and social groups, and are more sensitive to stylistic than to semantic variation.

How to participate

Upload your model to Hugging Face in a format compatible with STEB’s Hugging Face API. If you prefer not to release your model publicly, you can upload it to a private Hugging Face organization that only the shared-task evaluators can access. Detailed instructions, a starter kit, and the leaderboard will be linked here when the task opens.