Datasets:

syed-adbul
/

gridbench

id int64 0 6.42M	source_split stringclasses 2 values	clue stringlengths 1 130k	answer stringlengths 0 52
0	train	Carol's about to tolerate bad language	SWEARING
1	train	Explorer or actor Sebastian	CABOT
2	train	TV talk pioneer	PAAR
3	train	Utter failure, figuratively	LEADBALLOON
5	train	Denmark, a hyperactive place with a platform for free speech	HYDEPARK
6	train	Clean-up target	MESS
7	train	Archaic negative.	NILL
8	train	People slaloming	SKIERS
9	train	Man who runs a wine-shop	BODEGUERO
10	train	Late resident of Doorn.	KAISER
11	train	What's up, in Portuguese	EAI
13	train	Done, in poetry	OER
14	train	Bill Moyers speech on income inequality in America, with "The"	FIGHTOFOURLIVES
15	train	Started a film scene gradually	FADEDIN
16	train	Bitter salad veggie	ENDIVE
17	train	Unpopular insulation	ASBESTOS
18	train	Pearl container (wordplay)	OYSTER
19	train	Raft user	POLER
20	train	Brian ___.	BORU
21	train	Back to the Future destination	PAST
22	train	Rioting blamed for uproar	BEDLAM
25	train	More than simmering	ABOIL
26	train	House-warming method	GASHEAT
29	train	West of Broadway.	MAE
30	train	Thick-coated spitz type of dog originally from China	CHOW
31	train	Unaspirated wood tillers	ELMS
32	train	___ Sea (Italy/Greece separator)	IONIAN
33	train	Ham or beef	MEAT
34	train	Ukraine river	DONETS
35	train	Subatomic bit	ELECTRON
36	train	In a mournful manner	SADLY
38	train	Conductor's beginning (wordplay)	SEMI
39	train	Give a joint prize (wordplay)	REWARDPOT
40	train	Downloadable patch	BUGFIX
41	train	Foolery, sir, does walk about the ___ like the sun (Shakespeare)	ORB
42	train	Car with old sound	AUDIO
43	train	Will he at last bury his type of work (wordplay)	PRINTER
45	train	___ Paese (cheese choice)	BEL
46	train	Counting-out words for George	EENYMEANYMINEYMO
48	train	Potato sack race, essentially (wordplay)	HOPAGAINSTHOP
49	train	Cell fragment in the blood, important for clotting	PLATELET
50	train	Cultural matters, with "the"	ARTS
51	train	Shape of some docks	ELL
52	train	G.O.P. name.	NIXON
54	train	Cape, in place names.	NESS
55	train	Many an article by Ronnie on fruit	CITRON
56	train	Reportedly flog prisoner probably in here	CELL
57	train	Scantlings.	STUDS
58	train	Priscilla Mullens John	ALDEN
59	train	Important people	LIONS
60	train	Pool game with 21 balls	SNOOKER
61	train	Quickly Slang.	PDO
62	train	Huge format	IMAX
63	train	It comes in a small box	RING
64	train	Set in a groove	DADO
65	train	Boy present for a musical performance	SONATA
66	train	This bus might get one down before this	ERE
67	train	Medical drama locales abbreviation	ERS
68	train	Sixth-century date.	DXI
69	train	Google or Dropbox offering	FREESTORAGE
72	train	Sound made after breaking out (wordplay)	PEEP
74	train	Fish that yields caviare	STURGEON
75	train	Cheers Span.	OLES
76	train	___ in Calico: Crosby song	AGAL
77	train	Blazingly bright	AGLARE
78	train	Walter Reed rms.	ORS
79	train	Steel Ger	STAHL
80	train	Ex-N.J. Governor	KEAN
81	train	Maintain once more the RAs among the trees	REASSERT
83	train	Inspector's employer abbreviation	USDA
84	train	With a bow, on a score	ARCO
85	train	Lost no time	RACED
86	train	Mean persons	CAITIFFS
87	train	Pizza party leftovers	CRUSTS
88	train	Obscenities	FLETTERWORDS
89	train	Energetic one	DYNAMO
90	train	Texans' organization	NFL
91	train	Figures on the score board.	RUNS
92	train	It may not include a kiss	FIRSTDATE
93	train	Swampy forest region of Siberia.	URMAN
94	train	TV princess of Amphipolis	XENA
95	train	Divides equally	BISECTS
96	train	Former president of Romania	ILIESCU
97	train	Cast member of a play	STAGEACTOR
98	train	Controversial wiretapping group	NSA
99	train	The greatest follower of Buddhism initially in southeast Asia	BALI
101	train	Short ballet skirt	TUTU
103	train	Blackfish	TAUTOGS
104	train	Johnny a.k.a. Appleseed	CHAPMAN
105	train	Is in need of	LACKS
106	train	Actors.	TRAGEDIANS
107	train	Rare slip may produce an aggressive response	REPRISAL
108	train	Resource Conservation and Recovery Act organization	EPA
109	train	Newsroom measures	PICAS
110	train	Depp's 'Sleepy Hollow' co-star	RICCI
111	train	Irish operator is such a person in trade	SHOPKEEPER
112	train	Delta Tau Chi, familiarly	ANIMALHOUSE
113	train	Blow that for a cold spell!	SNEEZE
114	train	Body of moral values	ETHIC
115	train	Change over.	CONVERT

End of preview. Expand in Data Studio

GridBench

Crossword puzzles and clue-answer pairs for evaluating language models. The dataset has 98,548 puzzles, 5,332,845 entries, and 5,814,875 filtered CrosswordQA pairs.

Files

Config	Rows	Contents
`puzzles`	98,548	One row per puzzle with grid, block pattern, clues, answers, rebus metadata, and split.
`entries`	5,332,845	One row per clue-answer pair with puzzle ID, publisher, type, date, direction, coordinates, and split.
`crosswordqa`	5,814,875	Filtered CrosswordQA pairs with the original `source_split`.

Only the JSONL exports are here. Raw .puz files and raw CrosswordQA CSVs are not included.

Sources

The puzzles come from two sources.

The xwords archive is maintained by Q726kbXuN and backed by the q726kbxun.github.io repository. The archive stores puzzle data in xwords_data_00.dat through xwords_data_NN.dat, and its downloader is view_archive.py. The archive contains 94,225 JSON and text puzzles and 92,801 .puz files, dated 1932-05-15 to 2026-08-16.

The second source is xword-dl. This dataset uses the abdksyed/xword-dl fork at commit a586be8. The fork changes the Washington Post downloader from the Sunday endpoint to the daily endpoint. The GridBench download covers 13 xword-dl sources and produced 20,897 .puz files dated 2017-05-12 to 2026-08-22.

The Washington Post daily puzzle is the LA Times daily puzzle. WaPo copies are dropped in favor of the native LA Times file.

How the dataset was built

The build code is in the GridBench repository.

Download the xwords archive and the xword-dl sources.
Merge both sources and deduplicate by the SHA-256 hash of the solution grid. Archive files win over xword-dl files. Within xword-dl, LA Times wins over WaPo. This leaves 98,548 unique grids from 113,698 source files.
Classify the publisher from the source directory or xword-dl keyword. Classify the type from the filename suffix, then the source keyword, then the grid size.
Export the merged corpus to puzzles.jsonl and entries.jsonl with export_jsonl.py.
Assign splits to whole puzzles before expanding entries.
Filter CrosswordQA with export_jsonl.py.

The merge code is merge_data.py.

Grid size	Type
21x21	`sundays`
area up to 81	`mini`
area up to 169	`midi`
everything else	`classic`

Splits

Splits are assigned per puzzle, so every entry from a puzzle shares its split.

Split	Puzzles	Entries	Rule
`test_publisher`	8,188	280,779	Atlantic, Inkubator, New Yorker, Vulture, and Independent cryptic
`test_temporal`	9,036	407,557	Newest 10% of the remaining puzzles
`dev`	9,036	417,653	Next-newest 10%
`train`	72,288	4,226,856	Remainder

CrosswordQA

The source is CrosswordQA. Its train.csv and valid.csv are combined. A row is dropped when its normalized clue exactly matches a clue in any dev or test puzzle. Normalization lowercases the clue and replaces non-alphanumeric characters with spaces.

967,373 rows were dropped and 5,814,875 were kept. Exact clue collisions and exact clue-answer pair collisions with held-out puzzles are zero. Answer repetition is allowed by design.

License

License is not specified. Crossword content remains the property of its original publishers. Check the source terms before redistributing.

Downloads last month: -