robworks-software/jeopardy-clues

Jeopardy! Clue Dataset

A comprehensive, deduplicated Jeopardy! clue dataset — 568,068 unique clues spanning all 41 seasons (1984-2025), with 99.8% episode ID coverage across regular season, Tournament of Champions, Teen/Kids Tournaments, Teachers Tournaments, Celebrity Jeopardy, College Championships, Invitationals, and all special events.

Enriched with category analytics, repeat detection, difficulty estimates, NLP topic classification, and tournament/event metadata. Built by reconciling three sources (jwolle1, HuggingFace, J-Archive) with multi-pass deduplication and full provenance tracking.

Dataset Summary

Metric	Value
Total clues	568,068
Seasons covered	1-41 (1984-2025)
Unique episode dates	9,146
Episode ID coverage	99.8%
Unique categories	59,213
Topic-tagged clues	212,048 (37%)
Repeat clues detected	11,630
Daily Doubles	25,712+
Sources reconciled	3 (jwolle1, HuggingFace, J-Archive)

Special Event & Tournament Coverage

Event	Episodes
Tournament of Champions	778
College Championship	497
Teen Tournament	463
Celebrity Jeopardy!	249
Teachers Tournament	210
Jeopardy! Masters	157
Kids Week	109
Second Chance	99
Invitational	97
Battle of the Decades / Bay Area	64
Power Players Week	49
Million Dollar Celebrity	49
High School Reunion	30
Greatest of All Time	26
All-Star Games	19
Professors Tournament	16
Super Jeopardy!	14
IBM Watson Challenge	2
Unaired Trebek Pilots (1983-84)	2

Tournament and event metadata is available in the notes field for each clue.

Splits

Stratified random split (85/7.5/7.5), stratified by season so each split has proportional representation from all 41 seasons and the full date range (1984-2025):

Split	Examples	%
train	482,857	85.0%
validation	42,605	7.5%
test	42,606	7.5%

Features

Field	Type	Description
clue_id	string	Deterministic SHA-256 hash of composite key
air_date	string	Episode air date (YYYY-MM-DD)
season	int	Jeopardy! season number (1-41)
episode_id	int	Show number (99.8% coverage via J-Archive index)
round	string	jeopardy, double_jeopardy, final_jeopardy, or tiebreaker
category	string	Category name as displayed on the show
category_normalized	string	Lowercased, whitespace-normalized for grouping
value	int	Dollar value (null for Final Jeopardy / some Daily Doubles)
daily_double	bool	Whether this clue was a Daily Double
clue_text	string	The clue/prompt shown to contestants
answer	string	The correct response
clue_order	int	Position within category (1-5 for regular rounds, 1 for FJ)
category_frequency	int	How many episodes this category has appeared in
is_repeat_clue	bool	Whether a highly similar clue appeared in an earlier episode
repeat_clue_ids	list[string]	IDs of earlier similar clues (TF-IDF cosine >= 0.85)
topic_tags	list[string]	NLP-derived topic labels + difficulty:N (1-5 scale)
answer_word_count	int	Word count of answer
clue_word_count	int	Word count of clue text
notes	string	Tournament/event metadata (e.g. "2022 Tournament of Champions semifinal game 2")
sources	list[string]	Which source datasets provided this record
source_conflicts	string	Field-level disagreements between sources (JSON)

Topic Tags

Clues are classified into 17 topic categories using keyword and regex matching on both category names and clue text:

science biology space history war presidents geography us_states literature word_play movies television music sports food religion art

Each clue also has a difficulty:N tag (1=easiest, 5=hardest) estimated from dollar value, round, and Daily Double status.

Top 10 Categories

Category	Episode Appearances
science	355
american history	337
business & industry	323
literature	320
history	317
word origins	305
sports	298
world geography	296
potpourri	283
religion	262

Usage

from datasets import load_dataset

dataset = load_dataset("robworks-software/jeopardy-clues")

# Browse training data
print(f"Training clues: {len(dataset['train'])}")
print(dataset["train"][0])

# Filter by topic
science_clues = dataset["train"].filter(lambda x: "science" in x["topic_tags"])

# Get Final Jeopardy clues
final_jeopardy = dataset["train"].filter(lambda x: x["round"] == "final_jeopardy")

# Find hardest clues (difficulty 5)
hardest = dataset["train"].filter(lambda x: "difficulty:5" in x["topic_tags"])

# Filter Tournament of Champions clues
toc = dataset["train"].filter(lambda x: "Tournament of Champions" in (x["notes"] or ""))
print(f"ToC clues: {len(toc)}")

# Category frequency analysis
import collections
cats = collections.Counter(dataset["train"]["category_normalized"])
print("Most common categories:", cats.most_common(10))

Data Pipeline

1. Ingest: Three sources — jwolle1's GitHub release (529,939 regular + 21,592 kids/teen + 7,907 special events), openaccess-ai-collective/jeopardy from HuggingFace (216,930 clues), J-Archive scrape of incomplete episodes (5,745 clues)
2. Reconcile: Two-pass deduplication — composite-key matching across 3 sources, then text-level dedup to catch clues with identical text but different metadata across sources. 568,068 unique clues after removing 3,246 cross-source duplicates
3. Episode ID Mapping: J-Archive season index pages (41 seasons, 9,124 games) mapped show numbers to 99.8% of clues
4. Enrich: Category frequency stats (59,238 categories), TF-IDF repeat detection, value-based difficulty estimation, regex/keyword topic tagging (17 topics)
5. Export: Season-based train/validation/test splits as Parquet

Known Gaps

• 8 episodes with <50 clues: Mostly early-season episodes (1985-1989) and special formats (IBM Watson Challenge) where some clues weren't preserved
• Season 36: ~190 episodes due to COVID-19 production shutdown (March-June 2020) — not a data gap
• 0.2% missing episode IDs: ~1,200 clues (mainly from pre-season pilots and special events not in J-Archive's season index)
• clue_order defaults to 1: For many clues where position within category was not recorded in the source data

Use Cases

• Question answering: Train QA models on trivia across dozens of knowledge domains
• Text classification: Predict categories, difficulty, or topics from clue text
• Information retrieval: Build trivia search engines with category and difficulty filtering
• NLP research: Study question phrasing patterns, category evolution over 40+ years
• Educational tools: Generate quiz content across curriculum-aligned topics
• Analytics: Analyze trends in categories, difficulty, repeat patterns, and tournament structure over time
• Game AI: Build Jeopardy-playing agents with difficulty-aware strategy

Limitations

• Clue text and answers are the intellectual property of Jeopardy Productions, Inc.
• Some clues may have minor formatting differences between sources (tracked in source_conflicts)
• Topic tags use rule-based classification; ~63% of clues have no topic tag beyond difficulty
• Difficulty estimates are proxy-based (value tier, round) and do not reflect actual contestant performance
• Daily Double identification depends on source data quality

Legal Notice

Jeopardy! is a registered trademark of Jeopardy Productions, Inc. All question content is the property of Jeopardy Productions, Inc. and Sony Pictures Television.

This dataset is compiled from publicly available, community-maintained sources for non-commercial research and educational purposes under fair use (17 U.S.C. 107). The compilation, enrichment, and statistical analysis represent original work.

Sources & Attribution

• jwolle1/jeopardy_clue_dataset - Primary source (seasons 1-41, all tournaments and special events)
• openaccess-ai-collective/jeopardy - Secondary source for cross-validation
• J-Archive - Fan-maintained episode archive, used for episode ID mapping and gap-filling

This dataset should not be used for commercial purposes without appropriate licensing from Jeopardy Productions, Inc.