Gold Medal Page
Olympic-style medal tally across the v5 dataset cards. Each (dataset, target) is one competition. Only datasets with at least 5 criterion-model families competing count toward the medal table — uncontested entries (PsyProxy alone) are excluded. PsyProxy competes as one family (best of 4 lenses × strict/permissive + Sinhala translation arms). OpenAI competes as one family (best of Rathje-construct regression on GPT-4o-mini, GPT-4.1-nano, GPT-5-nano). Lexicon-based and topic-model baselines compete as their own singletons. Primary metric per task type: binary =FVE․Binomial, ordinal =Quadratic Kappa, regression =R², multilabel/multiclass =Macro F1. In the graph below, we represented these all as R²-equivalent scores.
Two things worth knowing before reading the graph. On the two leftmost tasks the ChatGPT wrapper makes almost no mistakes — AUC 0.995 on IMDb and 0.994 on Reddit suicidality. Both datasets come from public data-science competitions, so the relationship between those texts and those targets was most likely in the model’s training data. We read those two columns as a memorisation ceiling rather than a measurement of reading. Separately, the topic-model hyperparameters were tuned on the training folds: left at their defaults they returned R² near zero on almost every task.
Metric policy
Per-task podium
The same ten evaluations the animation walks through, in the same order, scored on the same strict holdout results. Each family enters its best member, so PsyProxy competes once — as whichever lens is strongest on that task. It takes first place on 7 of the ten.