MODEL_CARD.md

Model card: browser-compatible academic-rhetoric baselines

Task

Six-way text classification among background, objective, method, result, limitation and implication on an authored synthetic multilingual dataset.

Baselines

• Multinomial Naive Bayes (word features, additive smoothing).
• Word TF-IDF centroid cosine similarity.
• Character trigram TF-IDF centroid cosine similarity.
• Fixed 55% word / 45% character cosine hybrid.

Models are fit only on bundled 842 training examples. The application displays precomputed fully reproducible evaluations on 130 controlled and 54 novel-pattern examples. The user sandbox trains these baselines locally in the browser using the same code.

Results

At original 54-item challenge split: 74.1%, 72.2%, 70.4%, 77.8% accuracy respectively. Modified text perturbations affect the same examples, with no claim of verified semantic identity for every edit.

Risks and restrictions

Insufficient external validation, invented synthetic source population, uncalibrated score margins for similarity models, lack of human evaluation of meaning preservation, small subgroup counts, no demographic fairness audit, not intended for grading, hiring, admissions or safety-critical decision-making. No neural/foundation model training is claimed.