Jev Dimabsa

It never generates text. It asks described Score questions for valence and arousal and Choice and Noul questions for spans and pairs, then combines probabilities with corpus statistics and a frozen scorer.

Raw Jev scores sit on described rubric levels, so per-corpus ridge calibration maps them plus derived terms to the 1-9 gold scale. The penalty and design are chosen on train and dev, then frozen before test.

The Task 1 aggregate is the strongest reported result; Tasks 2 and 3 rank mid-pack because exact span boundaries decide most of the metric. Read the logs and per-corpus tables before assuming it fits your data.

Start with the quick-start, method overview, and experiment logs. You will need the dataset splits and the frozen example and calibration files, since train and dev choices are fixed before any test run.

Task 1 is the strongest: 1.0645 RMSE_VA on the official test across ten corpora, the lowest of the 14 teams with all ten, just below PAI at about 1.0663 and TeleAI at about 1.0737. The metric is the official scorer with normalization off, the root mean of squared valence and arousal errors over 16,186 gold entries, so lower is better and the margin over the best participant is 0.0018.

Per token, a Choice tags B, I or O; a lattice keeps every span the probabilities support plus train affix variants and literal vocabulary matches, then one Noul is asked per aspect-opinion pair. Four BM25-retrieved train reviews with annotated phrases show boundary conventions, the strongest single signal at AUC 0.80 to 0.90 and asked three times, and a per-language logistic regression reranks the pairs, lifting dev candidate coverage from 57% to 78%.

No. It is classification-only using Jev's Score, Choice and Noul, with dependencies limited to numpy and scipy on Python 3.12, and it runs Jev model 1.13.0 without any fine-tuning or GPU. What the pipeline adds instead is small corpus statistics: nine frozen train examples per corpus, a per-corpus ridge calibration, and a logistic regression reranker fitted on dev and frozen before the single official test run.