Spaces:
Sleeping
Sleeping
| 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 (0β100). | |
| """ | |
| 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`** β string. Use the image filename for VizWiz questions (e.g. `VizWiz_test_00002183.jpg`) and the numeric string ID for VQAv2 questions. | |
| - **`single_grounding`** β float between 0.0 and 1.0. Confidence that all answers share the same grounding region. `1` = single grounding, `0` = multiple groundings. | |
| """ | |