Baseline comparison
Four independently scored predictions: word Naive Bayes, word TF-IDF centroid, character TF-IDF centroid, and a weighted word/character hybrid. Scores are diagnostic, not safety ratings.
Generalization gap
A large IID-to-challenge gap indicates the limits of a controlled synthetic test set. The original Python model is archived separately, not presented as a head-to-head result.
Confusion matrix
Fixed evaluation partitions
Partitions come from the original Academic Rhetorical Intent Mini v1.1 dataset, not a freshly randomized split. Train rows never enter evaluation.
Reproduce, compare, record
Training is performed locally each time you run an experiment. The seed controls the bootstrap resampling, not the deterministic model fitting.
Bootstrapped uncertainty describes sampling variation in this specific synthetic evaluation set. It is not a model training or deployment uncertainty estimate.
Run a benchmark to see the results and 95% sample bootstrap intervals.
Run history
Up to 15 summaries are saved in this browser only. No raw academic text is stored in local storage.
Where each method succeeds or fails
Inspect performance by language, research domain, and rhetorical label. Small slices are descriptive only, not fairness certifications. Macro F1 averages labels that actually occur in the selected group.
Performance by language
Performance by domain
Per-label metrics
Inspect individual mistakes
Unlike a single score, the error explorer shows exactly what the method predicted and what the source label was.
Test your own sentence
Paste one academic-style sentence. No text is uploaded to a server. The displayed model scores are heuristic ranking signals, not calibrated probabilities.
Beyond the easy test score
Contrast controlled test outcomes against unseen patterns. This is a controlled synthetic shift analysis, not a real-world external evaluation.
Accuracy across evaluation settings
Archived Python classifier
These archived figures belong to a different Python TF-IDF classification pipeline and a distinct external biomedical sample. They come from the original project's evaluation results, not from this browser implementation. The 28-sentence sample is too small to establish general deployment performance.
Train/test integrity & provenance
The dataset was synthetically curated for this research portfolio. Audit checks make label coverage, partition sizes, and exact leakage signals visible.
Dataset composition
Leakage & integrity checks
Limitations and appropriate use
All examples in the original dataset are synthetic-curated. High accuracy can reflect repeated linguistic templates rather than general understanding. The challenge set is a harder fixed test; it is not a substitute for independent human-labeled real-world evaluation.
Do not use this system to grade students, evaluate academic integrity, screen admissions, determine employment outcomes, or certify scholarly quality.
Take every result with you
This project ships the entire evaluation dataset and executable JavaScript source. No hidden inference API, tokens, or remote server are required.
Export a reproducible record
Exports are generated locally, recording the evaluation split, method, seed, source version, and methodology notes.
Release methodology
- Use original fixed train / IID / challenge partitions.
- Fit each classifier to the same 842 training examples.
- Generate one prediction per held-out example.
- Evaluate Accuracy, Macro F1, confusion matrix and language/domain slices.
- Bootstrap accuracy over the chosen held-out sample with a fixed seed.
- Record data lineage and separate synthetic from real-world evidence.
Model governance
Data license: CC BY 4.0 (attributed to Academic Rhetorical Intent Mini). Code license: MIT.
Privacy: training and inference run entirely in the browser. History summaries, if enabled, remain in local browser storage.
Risk: These are interpretable baselines. The benchmark does not establish robustness outside controlled data or authorize consequential decisions about people.