--- license: cc-by-4.0 pretty_name: CADP-Bench task_categories: - image-to-text tags: - document-parsing - scientific-documents - latex - chart-to-code - multimodal - benchmark - arxiv:2608.17550 --- # CADP-Bench **A test benchmark for Compilable Academic Document Parsing.** CADP-Bench is a test benchmark for evaluating the reconstruction of academic document pages into contextual LaTeX and executable chart code. It covers continuous text together with four types of structured academic elements: **tables, mathematical equations, charts, and pseudocode**. The task emphasizes preserving document structure, surrounding context, and the ability to compile generated outputs back into a page for verification. Related paper: **[Code as Representation: A Compilable Parsing Paradigm for Academic Documents](https://arxiv.org/abs/2608.17550)**. [Paper](https://arxiv.org/abs/2608.17550) | [PDF](https://arxiv.org/pdf/2608.17550) | [Dataset repository](https://huggingface.co/datasets/seuai-wangjun/CADP-Bench) ## Benchmark overview The benchmark contains **1,630 page-level test samples** across **five academic domains**. Every sample in the index includes at least two structured element categories. Each sample is associated with a reference page PDF, a page-level LaTeX annotation, and a source directory containing the surrounding document context and supporting resources. The release occupies approximately **11.9 GB**, excluding local version-control metadata and caches. All samples constitute the benchmark's **test set** and are indexed in a single `data.json` file. ### Domain distribution | Domain | `subject` | Samples | | --- | --- | ---: | | Computer Science | `cs` | 1,047 | | Physics | `phy` | 251 | | Quantitative Biology | `bio` | 222 | | Statistics | `stat` | 65 | | Economics | `eco` | 45 | | **Total** | | **1,630** | ### Element counts and difficulty The following statistics are computed from the **current `data.json` annotations**. Element-level difficulty labels are **1 = Simple**, **2 = Medium**, and **3 = Hard**. Counts refer to individual elements, not pages; a page can contain multiple elements. | Element | Simple (1) | Medium (2) | Hard (3) | Total | | --- | ---: | ---: | ---: | ---: | | Table | 803 | 644 | 395 | 1,842 | | Equation | 340 | 536 | 147 | 1,023 | | Pseudocode | 40 | 71 | 27 | 138 | | Chart | 316 | 975 | 200 | 1,491 | ## Data construction As described in the paper, CADP-Bench is constructed from academic PDFs and their corresponding LaTeX sources collected from arXiv. The construction workflow includes source preparation and compilation checks, selection of pages containing multiple structured elements, alignment of page content with LaTeX annotations, chart reconstruction, and human verification. The retained document context supports evaluation of generated content within the original source environment, including document classes, packages, macros, references, and related assets. ## Repository structure ```text CADP-Bench/ |-- README.md |-- .gitattributes |-- data.json |-- pdf_data/ | `-- .pdf `-- data/ `-- / |-- main_flat.tex `-- ... source files and supporting resources ``` - **`data.json`**: a JSON object mapping sample IDs to annotations and metadata. - **`pdf_data/`**: 1,630 reference page PDFs, with paths specified by `page_path`. - **`data//`**: 1,630 sample directories containing document sources and supporting resources, such as figures, styles, bibliographies, and chart reference code where applicable. - **`main_flat.tex`**: the consolidated LaTeX source for the sample's document context. Sample IDs follow the pattern `_page`, for example `2201.04663v3_page11`. ## Annotation schema `data.json` is a **dictionary keyed by sample ID**, rather than a list of rows or a JSONL file. | Field | Type | Description | | --- | --- | --- | | `arxivid` | string | Source arXiv identifier, including its version. | | `page_num` | integer | Page number in the source document. | | `page_path` | string | Reference page PDF path, relative to the repository root. | | `latex_code` | string | Reference LaTeX annotation for the target page content. | | `image_paths` | list of strings | Referenced image paths, relative to `data//`. | | `insert_node` | integer | Stored insertion-position metadata for reconstructing content in the source context. | | `subject` | string | Academic domain: `cs`, `phy`, `bio`, `stat`, or `eco`. | | `type` | list of strings | Element categories present: `table`, `figure`, `equation`, and/or `algorithm`. | | `table_num`, `equation_num`, `algorithm_num` | integer | Counts of tables, equations, and pseudocode elements, respectively. | | `figure_num` | integer, optional | Number of chart elements. | | `table`, `equation`, `algorithm` | list of strings | Element-level reference LaTeX annotations. | | `table_scores`, `equation_scores`, `algorithm_scores` | list of integers | Element-level difficulty labels for the corresponding category. | | `figure_scores` | list of integers, optional | Chart difficulty labels. | | `layout_score` | integer, optional | Stored page-layout annotation. | The schema uses `figure` for charts and `algorithm` for pseudocode. Fields for absent elements may contain empty lists. In the current release, **342 samples omit `figure_num`, `figure_scores`, and `layout_score`**. Use `sample.get("figure_num", 0)` and `sample.get("figure_scores", [])` when reading chart annotations. The 1/2/3 difficulty mapping above applies to **element-level score lists**. Keep `layout_score` separate: it is optional and the current release also contains a value of 0. ## Getting started ### Download Install the download client: ```bash pip install -U huggingface_hub ``` Download the dataset into a local directory: ```python from huggingface_hub import snapshot_download snapshot_download( repo_id="seuai-wangjun/CADP-Bench", repo_type="dataset", local_dir="./CADP-Bench", ) ``` If you already have the dataset locally, skip this step. The download API is documented in the [Hugging Face download guide](https://huggingface.co/docs/huggingface_hub/guides/download). ### Read a sample Run the following from the parent directory of `CADP-Bench`, or adjust `root` to your dataset location. Reading the index uses only the Python standard library. ```python import json from pathlib import Path root = Path("./CADP-Bench") with (root / "data.json").open("r", encoding="utf-8") as f: samples = json.load(f) sample_id = "2201.04663v3_page11" sample = samples[sample_id] source_dir = root / "data" / sample_id page_pdf = root / sample["page_path"] source_tex = source_dir / "main_flat.tex" image_files = [source_dir / path for path in sample["image_paths"]] print("Samples:", len(samples)) print("Sample:", sample_id) print("Domain:", sample["subject"]) print("Element types:", sample["type"]) print("Reference page:", page_pdf) print("Source context:", source_tex) print("Referenced images:", image_files) print("Chart count:", sample.get("figure_num", 0)) print("Table difficulty:", sample["table_scores"]) print("Reference LaTeX preview:", sample["latex_code"][:300]) ``` ## Task and evaluation The CADP task takes a target page image as input and predicts contextual LaTeX for its text and structured elements, together with executable code for its charts. Reference page PDFs are provided in `pdf_data/`; render them into page images using the preprocessing configuration of your evaluation setup. The paper evaluates predictions using a **re-injection compilation** protocol: 1. Execute the model-predicted chart code to produce the required chart assets. 2. Insert the predicted LaTeX into the corresponding source context, with references to those assets. 3. Compile and render the reconstructed document, then compare the target page with the reference. The protocol measures executability and structural, textual, and visual fidelity. Reported metrics include execution rate, TEDS for tables, CDM for equations, Pseudocode Reconstruction Score (PRS), reading order, PageIoU, Visual Reconstruction Fidelity (VRF), and pixel similarity. The paper assigns a score of zero to samples that fail re-injection compilation. This repository provides the dataset assets and annotations. The examples above cover downloading and reading the release; they are not an end-to-end implementation of the paper's evaluation protocol. ## Usage notes - Preserve each sample's directory structure when resolving its referenced resources or compiling its source. - The consolidated source is document context, while `latex_code` is the target-page annotation; they serve different roles in reconstruction. - Some supplied chart reference scripts retain machine-specific output paths. Adapt those paths before executing the scripts on another machine. - Reading the provided pages, assets, and annotations does not require running the chart reference scripts. - The benchmark focuses on arXiv academic documents and pages with multiple structured element types; results should be interpreted within that scope. ## Citation If you use CADP-Bench in your research, please cite the paper using the [arXiv-provided BibTeX entry](https://arxiv.org/bibtex/2608.17550): ```bibtex @misc{jin2026coderepresentationcompilableparsing, title={Code as Representation: A Compilable Parsing Paradigm for Academic Documents}, author={Rihui Jin and Jun Wang and chengyuan zhu and Liang Mingyu and Yue Gao and Li Yunxuan and Kuicai Dong and Guilin Qi and Lin Ren and Yongrui Chen and Xinbang Dai and Jiaqi Li and Tongtong Wu and Gholamreza Haffari}, year={2026}, eprint={2608.17550}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2608.17550}, } ``` ## License The repository's declared license is **CC BY 4.0**, as specified in the dataset card metadata.