| --- |
| license: cc-by-nc-4.0 |
| pretty_name: VisualWorldBench-50K |
| --- |
| |
| # VisualWorldBench-50K |
|
|
| Training dataset for **VisualWorldBench: A Fine-Grained Multi-Task Benchmark for Evaluating Visual World Knowledge in MLLMs**. |
|
|
| **Authors:** Chen Chen, Zheming Liang, Jiaqi Wang. |
|
|
| [Project and code](https://github.com/chenchen0103/VisualWorldBench) · [Evaluation dataset](https://huggingface.co/datasets/orange0318/VisualWorldBench) |
|
|
| The release contains **7,670,789 training records** across MCQ, Open-QA, and grounding, covering **54,434 distinct target names**. |
|
|
| Compared with the data used in the reported experiments, this release includes additional source-level and annotation-quality filtering. |
|
|
| | Task | Complete annotation file | Records | |
| |---|---|---:| |
| | MCQ | `annotations/train.mcq.v3.jsonl.gz` | 6,850,706 | |
| | Open-QA | `annotations/train.open_qa.jsonl.gz` | 385,040 | |
| | Grounding | `annotations/train.grounding.jsonl.gz` | 435,043 | |
|
|
| The 47 MCQ `part-*` files are subsets covering the first 777,413 records of the complete MCQ file. For the full MCQ set, use `annotations/train.mcq.v3.jsonl.gz` alone. |
|
|
| Each compressed JSONL file contains one training record per line with `messages` and `images`. |
|
|
| MCQ and Open-QA records provide a direct image URL or an official dataset download page in `images`. Records with dataset-page URLs also include `path`, which locates the image within the extracted dataset. Obtain these images under the original providers' access conditions and replace `images` with local file paths before training. |
|
|
| Grounding records use image paths relative to the dataset directory. The **428,501 grounding images** are included in `image_shards/grounding-*.tar.gz`. Extract all grounding archives from the dataset directory: |
|
|
| ```bash |
| for archive in image_shards/grounding-*.tar.gz; do |
| tar -xzf "$archive" -C . |
| done |
| ``` |
|
|
| MCQ includes both ordinary and reasoning responses. Open-QA responses use `<think>...</think>` and `<answer>...</answer>`. Grounding responses contain bounding boxes in integer normalized `[x1, y1, x2, y2]` coordinates, as specified by each prompt. |
|
|
| Original training annotations are licensed under [CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/). This license applies to the annotations only; third-party images remain subject to their respective licenses and terms. |
|
|
| If you encounter any issues downloading the images, please contact the authors. |
|
|
