redasers/spelling-bee-human-difficulty

NYT Spelling Bee — Human Difficulty Dataset

Human solve-frequency data for 58 New York Times Spelling Bee puzzles (June–July 2025), sampled from 10,000 users per puzzle. Serves as ground truth for evaluating LLM orthographic constraint satisfaction.

Stat	Value
Puzzles	58
Date range	2025-06-02 to 2025-07-29
Total answer words	2,710
Words per puzzle	22–72 (mean 46.7)
Word length	4–13 characters
Users sampled per puzzle	10,000

Task

The NYT Spelling Bee presents 7 letters arranged in a honeycomb. Players must generate valid English words (4+ letters) using only those letters, with one designated center letter appearing in every word. Letters may be reused.

Schema

Column	Type	Description
date	string	Puzzle date (YYYY-MM-DD)
center_letter	string	Required center letter
outer_letters	list[str]	6 outer letters
answer_words	list[str]	All valid answer words
answer_user_counts	list[int]	Users (of 10,000) who found each word

Usage

from datasets import load_dataset

ds = load_dataset("redasers/spelling-bee-human-difficulty")
puzzle = ds["train"][0]

# Print puzzle setup
print(f"Date: {puzzle['date']}")
print(f"Letters: {puzzle['outer_letters']} (center: {puzzle['center_letter']})")

# Find the hardest and easiest words
words = puzzle["answer_words"]
counts = puzzle["answer_user_counts"]
pairs = sorted(zip(words, counts), key=lambda x: x[1])

print(f"\nHardest: {pairs[0][0]} ({pairs[0][1] / 10_000:.1%} solve rate)")
print(f"Easiest: {pairs[-1][0]} ({pairs[-1][1] / 10_000:.1%} solve rate)")

Data Source

Human solve frequencies were collected from the NYT Spelling Bee's public statistics, which report how many users found each answer word. A random sample of 10,000 users per puzzle provides the frequency counts.

Associated Paper

Bryan E. Tuck and Rakesh M. Verma. Orthographic Constraint Satisfaction and Human Difficulty Alignment in Large Language Models. Accepted at LREC 2026.

License

The puzzle structure and answer data are derived from the New York Times Spelling Bee game. This dataset is shared for non-commercial research purposes under fair use. The NYT retains all rights to the original game content.

Citation

@misc{tuck2025orthographicconstraintsatisfactionhuman,
      title={Orthographic Constraint Satisfaction and Human Difficulty Alignment in Large Language Models},
      author={Bryan E. Tuck and Rakesh M. Verma},
      year={2025},
      eprint={2511.21086},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2511.21086},
}