DATA_CARD.md

Data card: Synthetic rhetorical intent robustness records

Dataset form

Derived model evaluation snapshots with 4,416 prediction records from 184 source evaluation examples, 4 methods, and 6 conditions. The original fixed training/evaluation files are bundled for reproduction. Records have model method, source ID, split, language, domain, inherited rhetorical label, original and perturbed string, predictions, six model score values, top-two score margin and text-change flag.

Purpose

Research education on repeatable robustness evaluation and reporting of coverage. No natural language personal data is collected. Avoid using this dataset to evaluate real users.

Collection and transformations

Academic Rhetorical Intent Mini v1.1 examples were synthetic-curated. The stress-test transformations are deterministic JavaScript functions. Some synonym transformations leave examples unmodified because no dictionary match exists.

License

CC BY 4.0 for original example content and derived evaluation data; MIT license applies only to software. See DATA_LICENSE.md.

Limitations

Dataset is synthetic-curated and small. Inherited gold labels on edited texts are unverified, category prevalence is designed rather than natural, and test results cannot certify production reliability or robustness.