djdipin/data266-part3-inputs

DATA 266 Lab 2, Part 3: input images for the multi-agent system evaluation

DATA 266 Lab 2 (Fall 2026), Part 3, team 01 (Khushi Donda, Dipin Jassal). Code, frozen test cases and results: https://github.com/khushidonda/data266-lab2-team01 (src/part3/, evaluation/part3/).

These are the input images used by the 15 test scenarios and 10 red-team cases (role: eval), and the separate development images used only while building the system (role: dev). Each image fills a slot defined in evaluation/part3/scenarios.json / redteam.json; the slot rules and the selection are in evaluation/part3/build_assets.py, and every file's SHA-256 matches evaluation/part3/assets_manifest.json.

• 34 files (19 eval, 15 dev). metadata.jsonl gives, per file: slot, role, source, COCO id, SHA-256 and the licence conditions (attribution link and, where required, share-alike).
• Sources: COCO val2017 photos (Flickr) under licences that allow modification and redistribution; 8 of them carry synthetic printed text added by the team for prompt-injection red-team cases. Three files are deliberately broken inputs made by the team: S14_photo.png (the first 1024 bytes of a PNG, undecodable), R09_oversized.png (a blank 12000 x 12000 image) and R10_photo.png (a text file renamed to .png). The dataset viewer is off because of these.
• Not included: 6 photos of identifiable real people used in deceptive-edit red-team requests (slots R06, R07, R08 and their development counterparts). The team decided not to republish them.

Licences

Each photo keeps its own Flickr licence, listed with its attribution link in metadata.jsonl: CC BY 2.0, CC BY-SA 2.0, "no known copyright restrictions" or US government work. 10 files are CC BY-SA 2.0 and must be shared under that licence; images with printed text are adaptations of the original photo under the same licence. The COCO annotations are CC BY 4.0. The team's own synthetic files are free to use. For research and education.