Datasets:
| # TCA-Bench Evaluation Protocol | |
| The evaluator scores candidate audiovisual captions in four stages: | |
| - `1v`: Stage-1 visual caption quality against `gt/captions.json`. | |
| - `1a`: Stage-1 audio caption quality against `gt/captions.json`. | |
| - `2`: Stage-2 audio-visual binding accuracy against `gt/stage-2.json`. | |
| - `3`: Stage-3 temporal relation F1 against `gt/stage-3.json`. | |
| Prompt files are located in `prompts/`. `caption_prompt.txt` is the recommended prompt for generating candidate captions before evaluation. | |