Where G-Test Practice Questions Actually Come From: Six Sources Compared

Once the study method is settled, the next shortage is questions. Search for G-test past papers and, apart from the official twenty, you get sites listing items of unknown provenance. There are several problem sets, but which of them match Syllabus 2024? Do the courses include a mock exam? How far can you trust questions an AI wrote? With no map of where questions come from, the middle of your 40 days stalls.

This article contains affiliate links. As an Amazon Associate, we earn from qualifying purchases.

This article splits the sources into six, checks the volume, the quality of the explanations, the syllabus edition and the cost for each against primary material (the JDLA, the publishers, the course providers), and places each at a stage of study. The second half covers building a home environment that reproduces the real format, roughly 145 questions in 100 minutes, and handing the marking and the aggregation to an AI. All prices are as displayed officially on 21 September 2026, including tax.

First, what past papers actually are

G-test questions are not published as a rule. The only past items visible officially are the twenty the JDLA publishes as samples, drawn from sittings between the first of 2022 and the sixth of 2024 and labelled with the sitting and question number. Four options each, answers included.

So any resource lining up dozens or hundreds of items under the name past papers was not published by the examiner. It is either a reconstruction from candidate memory or a set of mock questions being called past papers. The former runs against the reason the exam is not published; the latter is a mock set under a different name. Either way there is no official backing for the answers, and the syllabus edition is usually old. Resources of that kind are not counted as a source here.

The premise for thinking about sources is the fact that there are only twenty official past questions. Everything else comes from the textbook, problem sets, courses, mock exams and AI generation.

The six sources compared

1. Official samples: twenty questions to learn the shape

The value of the samples is the shape, not the volume. Choose the single most appropriate option (12, including the expected-value calculation), the inverse, choose the single least appropriate (4, which is every one of the law and ethics samples), fill the blank (3) and choose the combination of terms (1): you learn what each of those looks like in practice. By major item, component techniques and applications get three each and the other eight get one or two, so all ten major items are represented.

Solve them once before you start and again in the middle. The second pass is to check whether your own glossary and your generated questions match this shape. Answers are published but explanations are not, so make yourself able to say why the right option is right.

One caution. Syllabus 2024 applied from the sixth sitting of November 2024, so only four of the twenty were set under the current syllabus; the remaining sixteen, from 2022, 2023 and the fourth sitting of 2024, come from the previous one. Check whether the term being examined survives into Syllabus 2024 using the keyword list. When this site matched the answer terms of all twenty against the keyword list of revision 1.4, nineteen were present, while one question on scaling laws from the fourth sitting of 2024 was not, and its stem contains ChatGPT, a term deleted in revision 1.4. Even official samples drift between the edition they were set under and the current one. You can check this for yourself, and the habit of checking builds the eye for choosing material by edition.

2. The official text, third edition: end-of-chapter questions

The third edition of the Deep Learning G-Test Official Text was published by Shoeisha on 27 May 2024, supervised by the JDLA, at 3,080 yen, A5, 448 pages. It conforms fully to Syllabus 2024, adds explanations of foundation models and large language models to every chapter, gives the law and ethics of AI its own chapter 8, and renews the end-of-chapter questions.

According to the JDLA introductory material, most people who pass work from two or three sources centred on the official textbook. The end-of-chapter questions are for solving right after reading the chapter; they are not where you build volume. Use them as a comprehension check and send the ones you get wrong back to your glossary and to mutual explanation.

If you feed a textbook you bought to an AI to generate questions, keep it within your own study. Sharing or publishing the generated questions or summaries goes beyond private use.

Explanations are another place AI helps. Problem-set explanations are short for reasons of space and often do not say why the other options are wrong. Hand the gist of the explanation and your wrong answer to the AI and ask which mid-level item and objective the explanation corresponds to, and why each distractor is wrong. Better than rereading the explanation is having the AI restate it and then restating it yourself. When the AI and the printed explanation disagree, go to the primary source.

3. Commercial problem sets: check the edition and the syllabus

The single most important thing when choosing a problem set is whether the edition matches Syllabus 2024. Anything published before the syllabus was released on 14 May 2024 is built on the previous one, and older editions contain terms such as ENIAC and fifth-generation computers that the current syllabus has deleted.

The JDLA recommended-reading page lists four titles under problem sets and exercises, but the edition shown there is not necessarily the newest. The Impress title, for instance, appears there as the second edition, while the publisher page shows a third edition (published 25 September 2024, 408 pages, 2,750 yen) written anew in line with the latest syllabus, formats and trends, with a final review set at the back matching the real exam in length. Use the recommended-reading page as an entry point and always confirm the current edition on the publisher site.

Autumn 2026 adds an option: Shoeisha, publisher of the official text, releases a companion problem set on 28 October 2026 at 2,750 yen with over 450 selected questions. That lands just before the November sitting, so if you are buying your volume book now, check the release date against your own plan. If you are already working, the third edition you can get today will move you faster.

How you use a problem set matters more than which one. Solve by major item and record your mistakes, and save the final review set at the back for the last fortnight. It is one of the few materials that supports a full-length run, and solving it early leaves you with nothing for the run-up.

Decide how you record, too. Think of a problem set as a tool for producing records rather than a tool for solving. Ticking the questions you have done means getting the same one wrong for the same reason on the second pass. Instead, for each mistake write four columns: major item, term, error type, note. Split error type into definition swap, wrong attribution (who, or which ministry) and missing condition (the rule and its exception), and the later aggregation shows you the kinds of weakness you have. On the second pass, solve only what is in that table. It takes less than half the time of redoing everything and it concentrates on the weak spots.

4. Courses: choose on the question service and the exercises

A course is an option for people who cannot sustain self-study, but it is worth looking at as a source of questions too. Two providers, checked on their own pages.

The Agaroot G-test course, revised for the 2024 No. 6 syllabus, costs 27,280 yen for about 9.5 hours of lectures. Alongside end-of-chapter questions it includes a final review set of 160 selected questions. The online question service can be used ten times during the period. Viewing access runs 180 days from the application date, and passing a G-test taken within that window earns a full refund of the fee, subject to conditions. Note that the official page states plainly that a mock exam is not included and that you should arrange your own.

The Study-AI G-test course costs 31,900 yen with about 12 hours of recorded video. The mock exam it calls Gmogi, a full set of more than 720 items, is included in the course, and the page states that the mock exam has been used by more than 40,000 people across more than 400 companies. The course outline lists the question-and-answer service as not provided.

When choosing, read the conditions on the official page in the original. The full refund carries a note that conditions apply, with specified sittings and application deadlines. A 180-day viewing window is ample for 40 days of study, but whether the deadline for the refund claim lines up with your sitting is yours to check. If the question service is capped at ten, do not spend it early; save it for the points the AI flagged in mutual explanation that you could not settle against a primary source.

5. Mock exams: a full run at the real length

A mock exam does two things: it puts the feel of 145 questions in 100 minutes into your body, and it shows you the unevenness of your scores across major items. Neither comes from a problem set or generated questions solved item by item.

Sources are limited. The final review set at the back of a problem set (the Impress third edition) and the mock included in a course (Study-AI Gmogi) are the main options; the Agaroot course does not include one. Free predicted-question sets exist, but where provenance and edition cannot be confirmed, judge them by the same standard as the sources above.

One or two runs in the last fortnight is enough. Rather than repeating, aggregate one result by major item and redirect the remaining days to the weak ones.

Read the score carefully. The JDLA does not publish the passing standard, so nobody can convert a mock score into a probability of passing. Instead of reacting to the total, look at the accuracy rate per major item. If one or two of the ten are clearly below the rest, that is where the remaining days go. If all ten are about level, the weakness is not an uneven distribution across items but the ability to separate pairs of terms, and you go back to mutual explanation.

6. AI generation: topping up weak mid-level items

The procedure and the measurements for having an AI write questions are in a separate article. With the syllabus supplied, one of 50 needed a correction; without it, twelve of 30, and the errors clustered on proper nouns in law and guidelines.

As a source, its place is topping up. Restrict it to the mid-level items your problem set or course does not cover, or the ones where mutual explanation exposed a hole, and generate five at a time. Because verification costs, it is not where you build volume. A generation run costs a few yen; verification costs a few minutes a question.

Generating questions for weak areas works best when the prompt pins the model to your material. This book covers prompt techniques for that kind of controlled generation.

Prompt Engineering for LLMs

USD 54.51 on Amazon.com (as of 2026/09/22)

Amazon

楽天市場

ポチップ

Placing the six across the stages

Total cost: self-study with the official text and a problem set is 5,830 yen; adding a course brings it to 33,110 yen (Agaroot) or 37,730 yen (Study-AI). Add the 13,200 yen exam fee and you are looking at just under 20,000 yen self-taught, or 46,000 to 51,000 yen with a course.

What you get for the paid route is exercises with explanations, a mock at the real length and, with a course, a question service. The mock in particular is hard to get for free. It changes how you spend the last fortnight of the 40 days, and buying a problem set purely for the mock at the back is a defensible decision.

Either way the role of the AI is the same. Fixing the scope, mutual explanation, topping up with generated questions and aggregating mistakes are all available whether or not you own the materials. What the materials supply is volume close to the real exam, not a way of studying.

Building the real format at home

Roughly 145 questions in 100 minutes cannot be reproduced by collecting questions alone. The environment includes the time constraint and the aggregation that follows.

Step one: assemble 145 questions. A mock or a final review set is the preferred material. Failing that, pull about 145 from the by-major-item questions in a problem set, weighted like the official samples, which lean towards component techniques and applications. If you mix in generated questions, use only the ones that passed verification.

Step two: match the duration and the hour. Set a timer for 100 minutes. Forty-one seconds a question is the pace; note the number of anything you skip and return at the end. Start at the same hour as your sitting if you can. The November sitting begins at 16:00 on Friday or 13:00 on Saturday, and concentration after lunch differs from concentration in the late afternoon. Do at least one full run in the slot you have booked.

Match the environment too. The recommended environment is published on the exam portal with version numbers, and tablets and smartphones are not recommended. Practise on a PC and turn off browser notifications.

Step three: hand the marking and the aggregation to an AI. When you finish, build a four-column table of question number, your answer, the correct answer and the major item. Give that to the AI and ask for the accuracy rate per major item and the three items with the most mistakes, and the aggregation takes seconds. Aggregation is work AI is good at: in a separate measurement on this site, the aggregation of a 21-row mistake log came back correct in all 16 cells.

The aggregate decides how the remaining days are split. The major items with low accuracy become the targets of next week mutual explanation and AI generation. Keep the record of mistakes rather than deleting it; it is also the input for the spaced repetition of the final fortnight.

Traps

First, stale editions. Bookshops still carry problem sets from the previous syllabus period. If the colophon predates November 2024, check on the publisher page whether it conforms to Syllabus 2024.

Second, unofficial past papers. Items of unknown provenance have no backing for their answers and are usually built on an old edition. Using material that runs against the reason the exam is unpublished is a way of devaluing the credential you are trying to earn.

Third, the quality of free predicted questions. Being free is not the problem. Sets with no explanations, or with options that are not drawn from the same category, do not train the shape. Compare against the official samples and judge on whether the shape matches.

Fourth, buying too much. The JDLA survey shows most people who pass get by on two or three sources. Three problem sets do not give you more time to solve. Working one by major item and accumulating your mistakes beats a single pass through three.

Summary

There are six sources, and the only official past questions are the twenty JDLA samples. Check chapter comprehension with the end-of-chapter questions in the official text (3,080 yen), build volume with a problem set that conforms to Syllabus 2024 (Impress third edition, 2,750 yen), and choose a course (Agaroot 27,280 yen, Study-AI 31,900 yen) on either the question service or the volume of its mock. Run a mock once or twice in the final stretch, and use AI generation to top up weak mid-level items.

Practise the real format with 145 questions in 100 minutes, at the same hour, on a PC. Hand the aggregation to an AI afterwards and push the major items with low accuracy into next week. Across 40 days, the mechanism that feeds results back into study matters more than the collecting of questions.

Sources

• JDLA, official sample questions (sittings and shapes of the twenty)
• JDLA, About the G-test (exam format, where to get the syllabus)
• JDLA, How to study for the G-test (survey on number of materials used)
• Shoeisha, Deep Learning G-Test Official Text, third edition
• Impress, Thorough Strategy Deep Learning G-Test Generalist Problem Set, third edition
• Agaroot, G-test preparation course (fee, lecture hours, question service, refund)
• Study-AI, G-test preparation course (fee, recording length, Gmogi, question service)

ブラウザだけでできる本格的なAI画像生成【ConoHa AI Canvas】

swiftwand

X

AIを使って、毎日の生活をもっと快適にするアイデアや将来像を発信しています。 初心者にもわかりやすく、すぐに取り入れられる実践的な情報をお届けします。 Sharing ideas and visions for a better daily life with AI. Practical tips that anyone can start using right away.