--- dataset_info: - config_name: css features: - name: structure dtype: string - name: text dtype: string - name: image dtype: image - name: download_url dtype: string - name: instance_name dtype: string - name: date dtype: string - name: additional_info dtype: string - name: date_scrapped dtype: string - name: file_filters dtype: string - name: compilation_info dtype: string - name: rendering_filters dtype: string - name: assets sequence: string - name: category dtype: string - name: uuid dtype: string - name: length dtype: string - name: difficulty dtype: string splits: - name: validation num_bytes: 815105541.0 num_examples: 300 download_size: 809865478 dataset_size: 815105541.0 - config_name: html features: - name: structure dtype: string - name: text dtype: string - name: image dtype: image - name: download_url dtype: string - name: instance_name dtype: string - name: date dtype: string - name: additional_info dtype: string - name: date_scrapped dtype: string - name: file_filters dtype: string - name: compilation_info dtype: string - name: rendering_filters dtype: string - name: assets sequence: string - name: category dtype: string - name: uuid dtype: string - name: length dtype: string - name: difficulty dtype: string splits: - name: validation num_bytes: 263470560.0 num_examples: 300 download_size: 257833986 dataset_size: 263470560.0 - config_name: javascript features: - name: structure dtype: string - name: text dtype: string - name: image dtype: image - name: download_url dtype: string - name: instance_name dtype: string - name: date dtype: string - name: additional_info dtype: string - name: date_scrapped dtype: string - name: file_filters dtype: string - name: compilation_info dtype: string - name: rendering_filters dtype: string - name: assets sequence: string - name: category dtype: string - name: uuid dtype: string - name: length dtype: string - name: difficulty dtype: string splits: - name: validation num_bytes: 279510653.0 num_examples: 300 download_size: 273214540 dataset_size: 279510653.0 - config_name: wild features: - name: image dtype: image - name: additional_info dtype: string - name: assets sequence: string - name: category dtype: string - name: uuid dtype: string - name: difficulty dtype: string splits: - name: validation num_bytes: 335841.0 num_examples: 2 download_size: 333134 dataset_size: 335841.0 - config_name: wild_legacy features: - name: structure dtype: string - name: image dtype: image - name: url dtype: string - name: instance_name dtype: string - name: date_scrapped dtype: string - name: uuid dtype: string - name: category dtype: string - name: additional_info dtype: string - name: assets sequence: string - name: difficulty dtype: string splits: - name: validation num_bytes: 99236852.0 num_examples: 50 download_size: 99142716 dataset_size: 99236852.0 configs: - config_name: css data_files: - split: validation path: css/validation-* - config_name: html data_files: - split: validation path: html/validation-* - config_name: javascript data_files: - split: validation path: javascript/validation-* - config_name: wild data_files: - split: validation path: wild/validation-* - config_name: wild_legacy data_files: - split: validation path: wild_legacy/validation-* --- # Image2Struct - Webpage [Paper](TODO) | [Website](https://crfm.stanford.edu/helm/image2structure/latest/) | Datasets ([Webpages](https://huggingface.co/datasets/stanford-crfm/i2s-webpage), [Latex](https://huggingface.co/datasets/stanford-crfm/i2s-latex), [Music sheets](https://huggingface.co/datasets/stanford-crfm/i2s-musicsheet)) | [Leaderboard](https://crfm.stanford.edu/helm/image2structure/latest/#/leaderboard) | [HELM repo](https://github.com/stanford-crfm/helm) | [Image2Struct repo](https://github.com/stanford-crfm/image2structure) **License:** [Apache License](http://www.apache.org/licenses/) Version 2.0, January 2004 ## Dataset description Image2struct is a benchmark for evaluating vision-language models in practical tasks of extracting structured information from images. This subdataset focuses on webpages. The model is given an image of the expected output with the prompt: ``` Please generate the source code to generate a webpage that looks like this image as much as feasibly possible. You should output a json object associating each file name with its content. Here is a simple example of the expected structure (that does not correspond to the image). In this example, 3 files are created: index.html, style.css and script.js. [ { "filename": "index.html", "content": "\\n\\n\\nTitle of the document\\n\\n\\n\\n

Content of the document......

\\n\\n\\n" }, { "filename": "style.css", "content": "body {\\n background-color: lightblue;\\n}\\nh1 {\\n color: white;\\n text-align: center;\\n}" }, { "filename": "script.js", "content": "document.getElementById(\\"demo\\").innerHTML = \\"Hello JavaScript!\\";" } ] You do not have to create files with the same names. Create as many files as you need, you can even use directories if necessary, they will be created for you automatically. Try to write some realistic code keeping in mind that it should look like the image as much as feasibly possible. ``` The dataset is divided into 4 categories. There are 3 categories that are collected automatically using the [Image2Struct repo](https://github.com/stanford-crfm/image2structure). The webpages were collected on GitHub pages (.github.io) and are split into 3 groups that are determined by the main language of the repository: * html * css * javascript The last category: **wild**, was collected by taking screenshots of popular websites. The full list is available at the end of this document. ## Uses To load the subset `html` of the dataset to be sent to the model under evaluation in Python: ```python import datasets datasets.load_dataset("stanford-crfm/i2s-webpage", "html", split="validation") ``` To evaluate a model on Image2Webpage (html) using [HELM](https://github.com/stanford-crfm/helm/), run the following command-line commands: ```sh pip install crfm-helm helm-run --run-entries image2webpage:subset=html,model=vlm --models-to-run google/gemini-pro-vision --suite my-suite-i2s --max-eval-instances 10 ``` You can also run the evaluation for only a specific `subset` and `difficulty`: ```sh helm-run --run-entries image2webpage:subset=html,difficulty=hard,model=vlm --models-to-run google/gemini-pro-vision --suite my-suite-i2s --max-eval-instances 10 ``` For more information on running Image2Struct using [HELM](https://github.com/stanford-crfm/helm/), refer to the [HELM documentation](https://crfm-helm.readthedocs.io/) and the article on [reproducing leaderboards](https://crfm-helm.readthedocs.io/en/latest/reproducing_leaderboards/). ## Citation **BibTeX:** ```tex @misc{roberts2024image2struct, title={Image2Struct: A Benchmark for Evaluating Vision-Language Models in Extracting Structured Information from Images}, author={Josselin Somerville Roberts and Tony Lee and Chi Heem Wong and Michihiro Yasunaga and Yifan Mai and Percy Liang}, year={2024}, eprint={TBD}, archivePrefix={arXiv}, primaryClass={TBD} } ``` ## List of websites used for wild subset ``` [ "https://www.nytimes.com", "https://www.bbc.com", "https://www.wikipedia.org", "https://www.github.com", "https://www.reddit.com", "https://www.twitter.com", "https://www.facebook.com", "https://www.instagram.com", "https://www.linkedin.com", "https://www.youtube.com", "https://www.amazon.com", "https://www.apple.com", "https://www.microsoft.com", "https://www.ibm.com", "https://www.google.com", "https://www.yahoo.com", "https://www.bing.com", "https://www.duckduckgo.com", "https://www.netflix.com", "https://www.hulu.com", "https://www.disneyplus.com", "https://www.imdb.com", "https://www.metacritic.com", "https://www.rottentomatoes.com", "https://www.nationalgeographic.com", "https://www.nasa.gov", "https://www.cnn.com", "https://www.foxnews.com", "https://www.bloomberg.com", "https://www.cnbc.com", "https://www.forbes.com", "https://www.businessinsider.com", "https://www.techcrunch.com", "https://www.engadget.com", "https://www.arstechnica.com", "https://www.lifehacker.com", "https://www.theguardian.com", "https://www.independent.co.uk", "https://www.buzzfeed.com", "https://www.vox.com", "https://www.theverge.com", "https://www.wired.com", "https://www.polygon.com", "https://www.gamespot.com", "https://www.kotaku.com", "https://www.twitch.tv", "https://www.netflix.com", "https://www.hbo.com", "https://www.showtime.com", "https://www.cbs.com", "https://www.abc.com", "https://www.nbc.com", "https://www.criterion.com", "https://www.imdb.com", "https://www.rottentomatoes.com", "https://www.metacritic.com", "https://www.pitchfork.com", "https://www.billboard.com", "https://www.rollingstone.com", "https://www.npr.org", "https://www.bbc.co.uk", "https://www.thetimes.co.uk", "https://www.telegraph.co.uk", "https://www.guardian.co.uk", "https://www.independent.co.uk", "https://www.economist.com", "https://www.ft.com", "https://www.wsj.com", "https://www.nature.com", "https://www.scientificamerican.com", "https://www.newscientist.com", "https://www.sciencedaily.com", "https://www.space.com", "https://www.livescience.com", "https://www.popsci.com", "https://www.healthline.com", "https://www.webmd.com", "https://www.mayoclinic.org", "https://www.nih.gov", "https://www.cdc.gov", "https://www.who.int", "https://www.un.org", "https://www.nationalgeographic.com", "https://www.worldreallife.org", "https://www.greenpeace.org", "https://www.nrdc.org", "https://www.sierraclub.org", "https://www.amnesty.org", "https://www.hrw.org", "https://www.icrc.org", "https://www.redcross.org", "https://www.unicef.org", "https://www.savethechildren.org", "https://www.doctorswithoutborders.org", "https://www.wikimedia.org", "https://www.archive.org", "https://www.opendemocracy.net", "https://www.projectgutenberg.org", "https://www.khanacademy.org", "https://www.codecademy.com", ] ```