Datasets:
| # Multimodal media | |
| | Archive | Role | Requests | | |
| | --- | --- | ---: | | |
| | [train.zip](train.zip) | Training / 训练集 | 4,000 | | |
| | [validation.zip](validation.zip) | Validation / 验证集 | 500 | | |
| | [standard_test.zip](standard_test.zip) | Standard test / 标准测试集 | 4,600 | | |
| | [robust_test.zip](robust_test.zip) | Robustness test / 鲁棒测试集 | 1,900 | | |
| Validation contains 400 standard and 100 robust requests for configuration | |
| selection. The test archives correspond directly to the technical report's | |
| Multimodal evaluation population. | |
| Images are optional when using the bundled CLIP features to fit or evaluate | |
| routers. Download the archives to inspect images, compute features, or run | |
| visual models on the benchmark requests. | |
| Each archive provides `instances.jsonl` with prompts and reference annotations, | |
| the required images, and a checksum manifest. Its `query_id` matches the Parquet | |
| data. A row's `media_json.archive` names the archive; `archive_member_path` or | |
| `archive_member_paths` identifies the corresponding image files. | |
| See [manifest.json](manifest.json) for archive checksums and | |
| [LICENSES.md](../../LICENSES.md) for upstream attribution and terms. | |