The mini split was generated with pandas' default NA parsing, which turned the literal answer string 'None' into an empty cell for one row. Regenerated as sample(n=300, random_state=42) of Infoseek_GDM_test with keep_default_na=False; this is the only changed cell.
Add FVQA, InfoSeek and EncyclopedicVQA evaluation splits (full + _mini)
Adds the three multimodal search benchmarks (FVQA, InfoSeek, EncyclopedicVQA), each as a full split and a `_mini` split (300 questions, `pandas.DataFrame.sample(n=300, random_state=42)` of the full split; `_mini` is the default evaluation split).
| Split | Rows | Image archive |
|---|---|---|
| FVQA_test | 1800 | images/FVQA.zip |
| FVQA_test_mini | 300 | images/FVQA.zip |
| Infoseek_GDM_test | 2000 | images/Infoseek_GDM.zip |
| Infoseek_GDM_test_mini | 300 | images/Infoseek_GDM.zip |
| EncyclopedicVQA_test | 1960 | images/EncyclopedicVQA.zip |
| EncyclopedicVQA_test_mini | 300 | images/EncyclopedicVQA.zip |
Same layout as the existing splits: `image_path` is relative to `images/<folder>/` inside `images/<folder>.zip` (stored, one top-level folder). All other TSV columns are byte-identical to the evaluation TSVs used in our runs, and every image was checked against its source file by sha256. README changes are additions only (configs, table rows, a note on `_mini`, and source/citatio…
Backup of processed counting/grounding splits (same as bryanzhou008/mini-vlm-toolkit-data)
