places_prompt_with_annotations_full_data.json

PLACES: Participatory Localized Red Teaming Dataset in the Global South for Text-to-Image Safety Description PLACES is a large-scale, community-driven red-teaming dataset for text-to-image (T2I) safety, collected through university partnerships across Ghana, Nigeria, Karnataka (India) and Punjab (India). Each record contains a prompt-image pair flagged by participants as harmful, together with structured harm/attack annotations and a free-text explanation of perceived failure. Intended use : evaluation and analysis of region- and culture-soecific failure modes in T2I systems, incuding cultural erasure, localized harms, and multilimngual/code-mixed prompt vulnerabilities. Dataset at a glance : Scale 26,139 prompt-image pairs (flagged harmful) 20,553 unique prompts 524 unique submitters (see note below on missing IDs) By locale (prompt-image pairs) Ghana : 6,565 Nigeria : 5,152 Karnataka (IN) : 7,536 Punjab (IN) : 6,886 Note on submitter IDs : due to a pricessing ussue, a subset of records may have missing submitter identifiers. Data format (JSONL) : The dataset is distibuted as JSON (and may also be provided as JAON Lines/ .jsonl: one JSON object per line). Each row corresponds to one submitte prompt and its associated generated image. Top-level fields (7 fields) : Field Type Description timestamp integer Unix timestamp indicating when the prompt was entered into Nibbler Date string (date format) Calendar date when the prompt was entered into Nibbler Prompt_id string The text of the prompt submitted into Nibbler submitted_prompt string The text of the prompt submitted into Nibbler hashed_filename string Unique numeric identifier (stored as a string) for the image associated with this record' used to link the record to the image file. Submitter_pool_id string Identifies which red-teaming challenge the submitter participated in (e.g., alpha, bravo, charlie, delta; if additional codes appear in the data, they correspond to additional challenges). Submitter_id integer An