answerdotai/ModernBERT-base, fine-tuned on SST-2

If you have ever called pipeline("sentiment-analysis") without naming a model, you have used distilbert-base-uncased-finetuned-sst-2-english. It is the default, it is downloaded millions of times a month, and it was fine-tuned on a backbone from 2019. This is the same task, on the same dataset, on an encoder released five years later.

The comparison is unusually clean, which is the reason this model is worth publishing at all. SST-2 has an official validation set of 872 examples that every SST-2 model reports on, so there is no split ambiguity: 0.9461 against 0.9130, a gap of 3.3 points, measured on the split both models are supposed to use. The card states the validation set was never used for training or checkpoint selection — a held-back 5% of train was used for that instead. SST-2 is also a saturated benchmark, so those 3.3 points are a real improvement on this data rather than a promise about yours.

Same two labels and the same semantics as the incumbent, so swapping the model string is the whole migration. What comes with those labels is the SST-2 convention: two classes, no neutral, English, sentence or short-paragraph length. Genuinely neutral text is forced to a side and comes back with a confident label, which is the thing to check before pointing this at survey or support data. It was trained on movie-review sentences and transfers well to general reviews and comments, less well to domain jargon such as financial text.