While preparing for the Greek Category B driving-theory test, I ran into a surprisingly basic problem: I could not find a complete English question bank that was easy to download, search, and study offline.
Practice websites exist, but they are not designed for inspecting the underlying data. I wanted a structured dataset, an illustrated PDF, and a clear record of which old answers might conflict with current Greek law.
This became a larger research and data-reconstruction project than I expected. ChatGPT Work was central to it—not just for coding, but for coordinating research, understanding an unfamiliar legacy database, auditing hundreds of questions, and rebuilding the study pack as new issues appeared.
The completed project is available here:
What ChatGPT Work contributed
ChatGPT Work helped locate and compare public sources, inspect the recovered database, write the extraction and PDF-generation code, and keep every output consistent after revisions.
More specifically, it:
- identified the most useful surviving source for the legacy English data;
- traced questions, answers, topics, languages, and illustrations across related records;
- compared the raw archive with a newer public question index;
- detected duplicate and superseded revisions;
- screened all retained questions for potentially outdated rules;
- researched current Greek legislation and safety guidance for flagged items;
- generated the JSON, CSV, audit files, ZIP, and illustrated PDF;
- built validation checks for counts, answers, duplicates, images, and stale currency;
- helped turn the work into a reproducible GitHub repository with documentation, source records, checksums, and licensing boundaries.
I still had to make judgment calls and manually check selected questions when public evidence was insufficient. ChatGPT Work connected those findings back into the workflow: a new problem could be researched, added to the structured audit, and reflected in a rebuilt PDF without manually editing hundreds of pages.
Reconstructing the legacy bank
The best starting point was the Greek Ministry’s old TestDrive package, released in 2009. A public extraction project had preserved its Microsoft Access data and illustrations.
ChatGPT Work helped inspect the program’s internal relationships. It isolated the English Category B records, connected questions to their choices and marked answers, restored topic information, and matched the image files. The first extraction contained 845 records.
That was not yet a clean question bank. The archive included older revisions and records no longer present in a newer public Category B question index. Comparing the two structures removed 11 no-longer-indexed records and resolved nine superseded revisions. The result was 825 unique questions with 414 illustrations and no exact duplicates.
Manual checks—and what they do not prove
There is no complete official public English question bank or question-by-question version history, so some uncertainties could not be resolved automatically.
I manually audited 56 questions across several chapters. After the legal review identified higher-risk topics, I checked six more involving child restraints, alcohol, insurance, accident wording, driving documents, and tyre replacement. They were consistent with the reconstructed records.
These checks establish consistency, not recency. They make me more confident that the reconstruction accurately represents the available legacy material. They do not prove that the questions were recently updated, that every marked answer reflects current law, or that the live examination uses exactly the same set.
That distinction became important because some faithfully recovered answers were clearly no longer suitable as present-day guidance.
Finding outdated or misleading material
ChatGPT Work screened all 825 retained questions and answer choices for time-sensitive content. The scan covered speed limits, alcohol thresholds, fines, emergency numbers, required documents, vehicle equipment, licence periods, parking distances, accident duties, ages, and other rule-dependent statements. Overlapping keyword and numeric checks produced 269 candidates for closer review.
Some conflicts were straightforward. The legacy bank gives 50 km/h as the general urban speed limit and uses 50 km/h as the motorway vehicle-capability threshold. Under the current Greek Traffic Code, the general urban limit is now 30 km/h, subject to specified exceptions, while motorway and expressway vehicles must be capable of exceeding 70 km/h.
Another question says that if only two new tyres are fitted, they should go on the front axle. That is genuinely the answer marked in the legacy data, but current tyre-safety guidance generally recommends putting the better pair on the rear axle to reduce the risk of losing rear-wheel grip.
The alcohol section exposed a different problem. The raw archive contained fine questions using drachmas, but those obsolete records were absent from the newer 825-question structure. The retained bank therefore has no drachma references, but also no replacement questions about current euro penalties. Rather than invent exam questions, the PDF includes a separately labelled guide to current alcohol rules.
In total, the final pack contains 14 source-backed legal or safety annotations, plus a visible warning about a first-aid requirement scheduled to take effect in 2027.
Preserving the old answer without teaching it as current
Silently replacing old answers would make the reconstructed dataset less faithful to its source. Leaving them unexplained could teach an outdated rule. The PDF therefore keeps two layers:
- Legacy bank answer: the choice marked in the recovered English TestDrive data.
- Current rule or safety note: a source-linked explanation where current law, translation, insurance treatment, emergency guidance, or safety advice differs.
This design was one of the most useful results of the iterative ChatGPT Work process. The project did not have to choose between being historically accurate and being useful for present-day study. It could preserve the bank’s answer while clearly showing why a learner should not treat every legacy statement as current guidance.
Making the result reproducible
The repository contains JSON and CSV data, all 414 illustrations, a 283-page PDF, audit reports, checksums, pinned dependencies, and the Python builder.
Running build_english_category_b_pack.py regenerates the dataset, ZIP, and PDF. Before producing the outputs, it validates the question count, checks that every question has one marked legacy answer, detects duplicate handling problems, confirms complete image coverage, and verifies that obsolete currency does not remain in the retained bank.
ChatGPT Work also helped make the builder portable instead of leaving it dependent on one machine. Font handling was updated for Windows, macOS, and Linux, paths were kept relative to the repository, and a clean rebuild was tested from the documented inputs.
The audit files record corrections, excluded obsolete records, manually checked groups, unresolved limitations, and modern topics the bank does not adequately cover. Each visible note can be traced to its record and cited source.
Sources, limitations, and rights
The legal review used the Hellenic Parliament record for Law 5209/2025, an article-level consolidated text, and the latest amendment checked at build time, Law 5322/2026, Article 76. Official sources also covered the 2024 textbook, emergency guidance, and examination rules. The dataset README contains the complete registry.
This remains an unofficial educational resource. It is not produced or endorsed by the Greek Ministry or any examination authority, may differ from the live examination, and may still contain errors or outdated material. It is not legal advice, professional driving instruction, or a guarantee of passing. The full limitations are in DISCLAIMER.md.
There is also an unresolved redistribution issue. The builder, project documentation, and original audit work use a limited-scope MIT licence. That licence does not cover the legacy questions, translations, illustrations, or source database; their provenance and uncertain rights status are explained in DATA_LICENSE.md.
The result is not a verified copy of the current live exam. It is a transparent and reproducible reconstruction of the publicly recoverable English legacy bank, with known conflicts clearly separated from the answers preserved by that bank. ChatGPT Work made it practical to move from scattered, aging material to a dataset that can be inspected, rebuilt, questioned, and improved.