{
  "version": "0.1.0-rc1",
  "working_title": "OTC-Bench: Overlay–Target Confusion Benchmark",
  "paper_title": "When Overlays Become Targets: Diagnosing Target–Distractor Confusion in Vision-Language Models",
  "code_repository": "https://github.com/Seren666/OTC-Bench",
  "data_repository": "https://huggingface.co/datasets/MortalSnow/OTC-Bench",
  "targets": 800,
  "pairs": 1600,
  "stimuli": 6400,
  "models": 6,
  "model_answers": 38400,
  "cohorts": {
    "old_96": 96,
    "old_204": 204,
    "new_500": 500
  },
  "source": "COCO val2014",
  "donor_instances": 16,
  "donor_categories": 8,
  "human_ratings": 3900,
  "human_reviewed_pairs": 1300,
  "cohort_semantics": {
    "old_96": "Earlier development/expansion cohort",
    "old_204": "Earlier extension, reviewed before new-500 freeze",
    "new_500": "Later frozen target validation with existing donors; no donor-instance holdout"
  },
  "generation": "See cohort-specific configs. Most runs use temperature 0/top_p 1/max_tokens 32; LLaVA uses do_sample false; some old96 seeds differ.",
  "public_image_redistribution": "not included; retrieve sources under their respective licenses and reconstruct",
  "scoring": "format_tolerant default (pre-existing uniform sensitivity) plus preserved frozen mode"
}