| license: cc-by-nc-sa-4.0 | |
| The NL-DIR dataset comprises 41K authentic document images, in which each image is paired with five high-quality fine-grained semantic queries, generated and evaluated through large language models in conjunction with manual verification. | |
| The complete dataset, along with detailed descriptions, specific formats, usage instructions, and construction methods, will be released soon. |