Academic Rhetoric Lab
Discover the rhetorical purpose of academic sentences, inspect the model's reasoning signals, and explore how evaluation changes beyond controlled examples.
What role does this sentence play?
Your text stays on this device. Scores are model outputs, not calibrated probabilities or factual judgments.
Predicted function
All category scores
Lexical signals
The word and phrase features that made the largest positive contributions to the predicted class in this compact browser classifier.
Six academic functions
Analyze a set of sentences
Paste one sentence per line. Each line will receive a predicted function, an uncalibrated model score, and a top-two margin. Great for exploring draft abstracts, not grading writing.
Predicted distribution
| # | Sentence | Prediction | Top score | Margin |
|---|
Explore the held-out challenge
These original, synthetic-curated sentences use unfamiliar templates compared with the training split. Filter by language or label to inspect where the compact browser model succeeds and fails.
| Language | Sentence | Reference | Prediction | Result |
|---|
Why one impressive score is not enough
We present favorable and unfavorable results side by side to show the difference between controlled testing and genuine domain transfer.
External biomedical recall by class
PubMed RCT: a 28-sentence development smoke test. Its classes required label mapping; it is not an untouched final external benchmark. The limitation category has no matching PubMed RCT class.
Reproducible assets
A complete research artifact includes a labeled dataset, model weights and training code, versioned evaluation reports, and a transparent live demo. The companion dataset and model can be published in separate Hugging Face repositories.