PHASES = [ {"label": "Dev (test-dev2025)", "codename": "test-dev2025"}, {"label": "Standard (test-standard2025)", "codename": "test-standard2025"}, {"label": "Challenge (test-challenge2025)", "codename": "test-challenge2025"}, ] LEADERBOARD_METRICS = [ "overall_f1", "overall_precision", "overall_recall", "vizwiz_f1", "vizwiz_precision", "vizwiz_recall", "vqav2_f1", "vqa_precision", "vqa_recall", ] DEFAULT_SORT_METRIC = "overall_f1" LEADERBOARD_FILE = "leaderboards/answer-therapy.jsonl" CHALLENGE_PHASE = "test-challenge2025" SUBFOLDER = "answer-therapy" CHALLENGE_TYPE = "VQA Answer Therapy" EVAL_DETAILS_MD = """ ### How is the Score Calculated? Each entry is a binary classification: does the visual question produce answers that all share the **same image region** (single grounding), or do different answers point to **different regions** (multiple groundings)? Your `single_grounding` confidence score is thresholded at **0.5** for evaluation. | Metric | Description | |--------|-------------| | `Overall F1` | F1 score across all questions *(default ranking metric)* | | `Overall Precision` | Precision across all questions | | `Overall Recall` | Recall across all questions | | `VizWiz F1` | F1 on questions from the VizWiz dataset | | `VQAv2 F1` | F1 on questions from the VQAv2 dataset | Scores are reported as percentages (0–100). """ FORMAT_MD = """ ### Submission Format Your JSON file must be a **list of result objects**, one per visual question: ```json [ { "question_id": "VizWiz_test_000000020000.jpg", "single_grounding": 0.85 }, { "question_id": "249549029", "single_grounding": 0.12 }, ... ] ``` - **`question_id`** — string. Use the image filename for VizWiz questions (e.g. `VizWiz_test_00002183.jpg`) and the numeric string ID for VQAv2 questions. - **`single_grounding`** — float between 0.0 and 1.0. Confidence that all answers share the same grounding region. `1` = single grounding, `0` = multiple groundings. """