CommonForms: A Large, Diverse Dataset for Form Field Detection

1. Introduction

• COMMONFORMS provides more than 480k pages from more than 59k document forms for form field detection.The paper presents the dataset as large, diverse, and high quality.Source: §1 · pp. 1–2
• The paper frames field detection as one step in form preparation, distinct from grouping fields and labels by their semantics.It focuses on visually predicting where fillable elements should go in a PDF.Source: §1 · p. 1
• More than 80 average precision on the COMMONFORMS test set, with higher recall and precision than Acrobat, is reported for FFDNet.The comparison with Acrobat is qualitative.Source: §1 · p. 1
• Cleaning Common Crawl PDFs improves data efficiency compared with using every PDF that contains a form field.The dataset is built from existing interactive PDF forms scraped from the Internet.Source: §1 · p. 1
• The detector predicts field locations and types from page images, including text inputs, signatures, and choice buttons.Choice buttons include checkboxes and radio buttons, which Adobe Acrobat and Apple Preview do not predict, as of publication.Source: §1 · pp. 1–2
• On 10k pages, varying input resolution produces a range of roughly 20 mean average precision points, underscoring the importance of high-resolution inputs.The FFDNet models use 1216px inputs.Source: §1 · p. 2

2. Related Work

• Form preparation includes making forms fillable, while form understanding extracts information from completed forms.The paper places its work in form preparation, described as the smaller of the two subfields.Source: §2 · p. 2
• Form2Seq [1] models semantic classification and grouping but assumes interactive widgets have already been detected; FormA11y uses human input to match labels to widgets.CommonForms instead provides a dataset and baseline models for automatic field detection.Source: §2 · p. 2
• Form understanding methods extract key-value pairs from completed forms, with datasets such as FUNSD [10] and NAF [5] containing 199 and 77 labeled forms, respectively.Methods include visual, textual, multimodal, and graph-based approaches.Source: §2 · pp. 2–3
• CommonForms treats documents as unimodal images and uses object detectors to localize and type form fields, producing a semantic layer complementary to other document models.The paper distinguishes this visual prediction of value slots from extracting values themselves.Source: §2 · p. 3
• Like ccPDF [22] and FinePDFs, CommonForms draws from Common Crawl, but its filtering targets well-annotated forms for field-detection training.Source: §2 · p. 3

3. COMMONFORMS Dataset

• The authors filter an approximately 8-million-PDF Common Crawl sample to mine well-prepared forms as training data, without manual annotation.They argue that existing fillable forms provide a training signal for bootstrapping a detector.Source: §3 · p. 3

3.1. Dataset Preparation and Cleaning

• Filtering 7.9 million PDFs for form objects leaves 762k; removing PDFs without fields or with only Button fields leaves 59k, and field cleanup yields 480k pages.The document-level split reserves 8k validation pages and 25k test pages to prevent similar pages from leaking across splits.Source: §3.1 · pp. 3, 5

3.2. Annotation Consistency

• Despite filtering out more than 99% of candidate documents, forms retain inconsistencies such as spurious fields, signature mismatches, and semantically ambiguous sections.The authors leave these inconsistencies unfiltered as a trade-off for scaling the dataset; qualitative results still find the models useful on complex forms.Source: §3.2 · p. 5

3.3. Language Identification

• FastText classifies page languages from extracted text; English accounts for 63.6%, while the remaining third includes languages such as Cantonese, German, Korean, Spanish, and French.Textual cues and differences in right-to-left versus left-to-right field placement make examples across languages useful for training.Source: §3.3 · pp. 5–6

3.4. Domain Classification

• LDA topic modeling produces 300 topics that are labeled into 14 domains, with mixed or unclear topics assigned to Other.Outside Other, the five most common domains are Government and Administrative, Commerce and Tax, Engineering, Data and Privacy, and Law and Justice.Source: §3.4 · p. 6

4. FFDNet

• FFDNet-Small and FFDNet-Large are YOLO11 detectors initialized from scratch, with 9 million and 25 million parameters, respectively, for three form-field classes.The classes are Choice Button, Text Input, and Signature.Source: §4 · p. 6
• Both models use 1216px inputs, a resolution chosen to balance training and inference efficiency against performance.The paper says later experiments support the necessity of high-resolution models; larger resolutions reduce batch sizes and lengthen training.Source: §4 · p. 6
• Both models were trained for 300 epochs with an initial learning rate of 0.001 on 4xV100 instances; FFDNet-L took roughly 5 days to train, while FFDNet-S took roughly 2 days.Source: §4 · p. 6
• The authors chose hyperparameters and input resolution by training several models on subsets of CommonForms and examining generalization as model and dataset sizes scaled.Source: §4 · p. 6
• A processing error meant both models were trained on approximately 350k form pages rather than the full 490k pages.The authors say they are working on retraining the models for release.Source: §4 · p. 6

5. Experiments and Results

• FFDNet-L consistently outperforms FFDNet-S across all classes.On a single 3090Ti, FFDNet-L takes roughly 16 ms per page versus roughly 5 ms for FFDNet-S.Source: §5 · p. 6, Table 3 · p. 8

5.1. Resolution Matters In Form Field Detection

• About 20 performance points separate 640px and 1536px inputs; Choice Button and Signature detection are most affected.Small widgets and weak textual cues such as signature indicators help explain the sensitivity; 1216px is described as a speed–accuracy trade-off.Source: §5.1 · p. 6

5.2. FFDNet Outperforms Adobe Acrobat at All Sizes

• Acrobat does not detect choice buttons and has substantially lower precision and recall than FFDNet for text and signature fields.The comparison is qualitative.Source: §5.2 · p. 6, Table 2 · p. 7

5.3. FFDNet is Robust Across Languages/Domains

• Both models perform similarly across 9 of the 10 most common languages but degrade in Russian; both are consistent across domains, with FFDNet-S more variable.Source: §5.3 · p. 8

5.4. Filtering Improves Data Efficiency

• mAP50−95 was 57.9 versus 53.6 for 10k pages sampled from the filtered 59k-document set versus the all-form 760k-document set.Source: §5.4 · p. 8

• COMMONFORMS is a released dataset of 490k diverse form images filtered from PDFs in Common Crawl.Source: §6 · p. 8
• The authors release FFDNet-S and FFDNet-L, which qualitatively outperform Adobe Acrobat, together with the dataset and preparation code.Source: §6 · p. 8
• Future work targets semantic form preparation and improvements for scans, foreign-language documents, and form-preparation inconsistencies.Source: §6 · p. 8