| --- |
| license: mit |
| pretty_name: MIB Doc Challenge Data |
| task_categories: |
| - image-to-text |
| - text-classification |
| language: |
| - en |
| tags: |
| - document-processing |
| - ocr |
| - synthetic |
| size_categories: |
| - 1K<n<10K |
| --- |
| |
| # MIB Doc Challenge Data |
|
|
| Synthetic PDF packets for 8090's MIB Doc Challenge: Intergalactic Intake. Participants extract structured fields from messy, adversarial document packets and classify each case as `APPROVED`, `DENIED`, or `NEEDS_REVIEW`. |
|
|
| ## Contents |
|
|
| The versioned archive `mib-doc-challenge-public-data-v2026-07-07.zip` contains: |
|
|
| - `data/train/`: 1,000 training PDFs |
| - `data/train_labels.csv`: public labels for the training PDFs |
| - `data/validation/`: 5,000 unlabeled validation PDFs |
| - `data/validation_manifest.csv`: validation case IDs, paths, and page counts |
|
|
| Validation answers and the final private test set are not included. |
|
|
| ## Integrity |
|
|
| SHA-256: |
|
|
| ```text |
| a9bb8c1bbf51346ebf49c2e3e1acdb7a5d6cd0760162767b0d133c7b7200f3c4 mib-doc-challenge-public-data-v2026-07-07.zip |
| ``` |
|
|
| ## Data characteristics |
|
|
| All applicant identities, documents, portraits, and case records are synthetic. The PDFs intentionally include scan degradation, conflicting evidence, hidden text, fake answer keys, and other adversarial content. Hidden instructions are part of the challenge and are not trustworthy labels. |
|
|
| ## Usage |
|
|
| Use this dataset with the public challenge repository at <https://github.com/8090-inc/mib-doc-challenge>. The repository contains the field manual, schemas, evaluator, submission validator, and Docker runtime contract. |
|
|
| ## License |
|
|
| Released under the MIT License. |
|
|
