| # E2 Benchmark |
|
|
| An efficiency benchmark of 180 VQA cases in two sweeps. Each sweep varies one workload dimension |
| while holding the others fixed, so latency can be measured against that axis. Token counts use |
| LLaVA-1.5's tokenizer and prompt format. Images are unmodified, so visual tokens are constant only for |
| fixed-resolution models. |
|
|
| | `category` | varied axis (`level`) | levels | cases | held fixed | |
| |---------------------------|------------------------------|---------------|-------|------------------------------------| |
| | `input_text_tokens` | prompt tokens | 16 (57 to 74) | 80 | answer 1 to 11 tokens | |
| | `reference_answer_tokens` | gold answer tokens | 20 (1 to 21) | 100 | prompt 59 to 63 tokens | |
|
|
| Each level has 5 distinct cases. |
|
|
| ## Sources |
|
|
| | dataset | split | cases | |
| |---------|------------------|-------| |
| | AI2D | test | 8 | |
| | ChartQA | test (human) | 37 | |
| | DocVQA | val | 65 | |
| | GQA | testdev balanced | 23 | |
| | TextVQA | val | 47 | |
|
|
| AI2D questions that refer to diagram letters use the ABC-labelled version of the diagram (`images/AI2D_abc/`). |
| AI2D answers are given as option text, not letters. |
|
|
| ## Layout |
|
|
| ``` |
| varest_benchmark.json # all cases |
| images/<dataset>/ # source images |
| ``` |
|
|
| ## Fields |
|
|
| - `question_id`: unique case id |
| - `image`: image path, relative to this folder |
| - `text`: question |
| - `answer`, `answers`: primary gold answer, all gold answers |
| - `category`, `level`: sweep and its value for this case |
| - `evaluator`: scoring method (ANLS for DocVQA, relaxed accuracy for ChartQA, VQA soft accuracy for TextVQA, |
| exact match for GQA, option match for AI2D) |
| - `source`: original dataset, split, question id and file paths within the original dataset release |
| - `workload`: visual, prompt and answer token counts, and image size |
|
|
