Bubsy AI subscriber data - stratified sample
Description
Compressed in a zip file, the dataset accompanying this work is a 10% stratified, de-identified extract of the moderation log, comprising 48K messages drawn from 27 communities active between November 2024 and July 2025. It is distributed as two flat CSV files, retaining the table and column names of the source log so that the analysis scripts run against it unchanged. Sampling is stratified on the fields the reported metrics are computed from: actions on the outcome of the action (missed, reverted or clean), the community and the triggering detector; messages on the processing path (text, image or short-circuited), the timing state and the community. Each stratum contributes its own 10%, because a single global draw loses the rare cells that carry the error rates (the 22 reverted mute actions in particular). Messages referenced by a sampled action are retained, with an unreferenced message from the same cell dropped in exchange, so the stratum counts and the latency mix are preserved. Community names, usernames, target user identifiers and message identifiers are replaced with randomly generated UUIDs under a one-to-one mapping that is not retained; free-text detector explanations are removed, as they can quote the message itself, and the term lists are reduced to a present or absent marker. The extract reproduces the reported precision (0.99838 against 0.99822) and recall (0.97476 against 0.97484), and the mean, median and 90th-percentile processing times to within 1.2%.
Files
Institutions
- Karachi Institute of Economics and TechnologySindh, Karachi