Sensible
Sensible defaults for everyday NLP. Modern architectures, honest evaluation, and models you can drop into production without a research detour.
Models
Sensible ModernBERT Sentiment
SST-2 accuracy 0.9461 vs the incumbent default's 0.9130, measured on the official validation set.
Benchmarks
SST-2 official validation set, the split every SST-2 model reports on, never used for training or checkpoint selection:
| Model | Accuracy |
|---|---|
| Sensible ModernBERT | 0.9461 |
| distilbert-sst2 (the 3.9M-downloads/month default) | 0.9130 |
Training script and raw eval outputs ship in the repo.
Use it
from transformers import pipeline
clf = pipeline(
"text-classification",
model="AnkitAI/Sensible-ModernBERT-Sentiment-Analysis",
)
clf("This movie was absolutely wonderful!")
# [{'label': 'positive', 'score': 0.99}]Roadmap
More everyday classifiers are coming: the same recipe of modern architectures and honest evals, applied to the text tasks teams reach for most.
Reproducibility
The reported number comes from the official SST-2 validation set, which was never used for training or checkpoint selection. The training script and raw evaluation outputs ship in the model repo, so the result can be re-run end to end.