I have been looking at trivia game stuff recently and also been thinking about ROM emulation files, for old console gaming systems like Nintendo, as a data corpus. And it made me remember seeing these Jeopardy games around growing up and thinking how awful they must be. But now that I’m old I kind of think they are neat, or at least conceptually. There ended up being quite a lot of games released over the ~10 years that cartridge systems were hugely popular. And I wondered what kind of questions were on these things? The market is of course mostly for children but the Jeopardy brand is pretty serious trivia. So I gathered them up and extracted the questions and did some analysis.

Getting the questions out I made use of a coding agent to search the binary ROM file. This is a good example of using them where it can just churn away brute forcing binary offsets of the 14 files for as long as needed. I did find a hint before I started that the early games used the classic Huffman coding algorithm so that was a good start:

You can see the early games were space restricted so there are more elaborate approaches to packing the questions into the cartridge. While at the end of the run, the tens of megabytes the N64 allows for storage enabled luxurious plain ASCII text.

There were 34K questions extracted total but if you remove duplicates across games you end up with 24,239 unique questions. That’s a lot of questions! I wanted to get a sense of what kind of questions appeared in each game. So I did a process that is becoming my goto workflow of embedding -> clustering -> sample cluster members. I used Google’s embedding 2 model, did HDBSCAN and k-means clustering and used my local Qwen3.6-35B-A3B model to sample the members and categorize them. I UMAP’d them into this interface:

Note: I do not make the questions dataset available for download anywhere, they are in this visualization purely for informational purposes, see the J! Archive for Jeopardy trivia questions.

If we remove the sports themed games and use the more broad k-means clusters these are the top 10 categories:

Rank	Category	Questions
1	Animals	733
2	U.S. Presidents	710
3	Idioms and Slang	691
4	Countries and Capitals	623
5	General Knowledge Definitions	601
6	Games And Toys	591
7	Classic Television Shows	562
8	Classic Hollywood Films	561
9	US Cities	551
10	American Founding Figures	548

There is of course the "Junior" version of the game that has more kid friendly questions, but the majority questions I would say are geared towards adults. There are a few areas of overrepresentation by game/platform. For example the Sega Genesis version has over two times the amount of biblical questions compared to other versions. And the NES 25th Anniversary version has four times as many questions about The Beatles than any other version. Other areas of over-representation using the HDBSCAN clusters:

Game	Over-represented category	Lift	Count
Sega Genesis (v1)	Biblical Figures & Texts	2.3x	68
Sega Genesis Deluxe	Gems and Birthstones	4.6x	30
Sega Genesis Deluxe	Ancient Egypt	3.1x	33
Sega Genesis Deluxe	Fabrics and Fashion	2.6x	73
N64	Caribbean Geography	3.7x	23
N64	Scandinavian Nations	3.6x	31
N64	Middle East Geography and History	3.1x	44
NES 25th Anniversary	The Beatles	4.0x	14
NES 25th Anniversary	Country Music Artists	2.5x	87
Super Jeopardy!	Classic TV Sitcoms	2.3x	86
SNES (v1)	American Revolution	2.8x	34
Game Gear / Game Boy	Magazines and Newspapers	2.8x	16

And there are a few questions that repeated across many of the game's iterations:

Games	Clue	Answer	Jeopardy Category
5	THE FASTEST GAIT OF A HORSE	GALLOP	ANIMALS / HORSE RACING
4	HE WAS PROCLAIMED "ADMIRAL OF THE OCEAN SEA" BY FERDINAND & ISABELLA	CHRISTOPHER COLUMBUS	ADMIRALS
4	FRENCH EMPEROR WHO REDESIGNED THE ITALIAN FLAG TO RESEMBLE HIS OWN	NAPOLEON	ITALY
4	WEST INDIAN ISLAND THAT'S THE HOME OF THE LIMBO DANCE & CALYPSO MUSIC	TRINIDAD	ISLANDS
4	WHETHER ST. PATRICK WAS RESPONSIBLE OR NOT, NONE ARE NATURALLY FOUND IN IRELAND	SNAKES	ANIMALS
4	SAUTEED EGG-DIPPED BREAD	FRENCH TOAST	TOAST
4	A CAMPFIRE CONFECTION	TOASTED MARSHMALLOW	TOAST
4	"RANK" OF GEORGE JESSEL	TOASTMASTER GENERAL	TOAST
4	NAMED FOR AN AUSTRALIAN SOPRANO, IT'S THIN, CRISP BREAD	MELBA TOAST	TOAST
4	THIS NEIGHBORING COUNTRY'S FRANC IS WORTH MORE THAN FRANCE'S	SWITZERLAND	FOREIGN EXCHANGE
4	ANTI-MALARIAL DRUG PRESENT IN TONIC WATER	QUININE	MEDICINE

Finally I was thinking, out of 24K questions there HAD to be some wrong ones in there. I don’t mean question answer pairs that were correct in 1989 but wrong today I mean straight up incorrect. To find these I used a local LLM Qwen 3.6 35B A3B to scan 20 questions at a time to flag possible issues. The second pass then did a more thorough skeptical check that thinned down that previous pass and then finally sent those that kicked up as wrong in both passes to gemini-3.6-flash with google search grounding enabled. Then I reviewed the output of that last pass and found over 50 questions that were incorrect. Oftentimes the wrong questions are just an issue of the writer inverting the clue or other minor problem but sometimes boldly incorrect. Here is a list:

It’s funny to think of some 1993 teen playing Jeopardy on their SNES and being convinced for years afterward that Plessy v. Ferguson actually banned segregation or that the Gregorian calendar started in 1852. You can browse the screenshots here to get a sense of what these games looked like and there are a few playthroughs on Youtube if you want to torture yourself.

Source code on GitHub.