floleuerer/strawberry-bench

StrawberryBench Dataset

Dataset Description

StrawberryBench is a benchmark dataset for evaluating the ability of Large Language Models to count letter occurrences in words and phrases. The task appears trivial but exposes a fundamental limitation: LLMs operate on sub-word tokens, not individual characters, making direct character-level counting non-trivial.

The canonical example: "How many r's are in 'strawberry'?" (answer: 3). This question went viral when ChatGPT famously answered incorrectly, revealing a systematic weakness tied to Byte-Pair Encoding (BPE) tokenization.

Supported Tasks

• letter-counting: Given a word or phrase and a target letter, count how many times the letter appears.

Languages

English (en)

Dataset Structure

Data Instances

{
  "id": "sb_00001",
  "word": "cat",
  "letter": "t",
  "question": "In the word 'cat', how many 't's are there?",
  "answer": 1,
  "difficulty": "easy",
  "word_length": 3,
  "zero_count": false,
  "template_idx": 1
}

Data Fields

Field	Type	Description
id	string	Unique example identifier (sb_XXXXX)
word	string	The word or phrase to count in
letter	string	The single letter to count (lowercase)
question	string	Natural language question
answer	int	Correct count (ground truth)
difficulty	string	easy / medium / hard / sentence / paragraph / names / foreign
word_length	int	Number of non-space characters in word
zero_count	bool	Whether answer == 0 (letter absent)
template_idx	int	Which question template was used (0 or 1)
language	string	For foreign difficulty, the language (optional)

Data Splits

Split	Examples
train	—
test	847

The dataset is released as a single test split only; it is designed for zero-shot / few-shot evaluation, not fine-tuning.

Difficulty Tiers

Tier	Description / Word Length	# Examples
easy	3–6 chars	183
medium	7–12 chars	228
hard	13+ chars	146
sentence	short multi-word phrases	46
paragraph	longer text passages (100+ chars)	55
names	common first names (repeated letters)	112
foreign	non-English words (German, etc.)	77

Approximately 11% of examples have zero_count = true (the letter does not appear in the word).

Dataset Creation

Source Data

All words and phrases are common English vocabulary, proper nouns, and well-known idioms. No copyrighted text is included. Ground-truth answers are computed deterministically using Python str.count().

Question Templates

Two main templates are used per example (randomly assigned index 0 or 1):

• Standard (easy, medium, hard):
  1. "How many times does the letter '{letter}' appear in the word '{word}'?"
  2. "In the word '{word}', how many '{letter}'s are there?"
• Sentences: Uses "phrase" instead of "word".
• Paragraphs: Uses "following text: '{word}'".
• Names: "How many {letter}'s are in the name '{word}'?"
• Foreign: "How many times does the letter '{letter}' appear in the {language} word '{word}'?"

Curation Rationale

The dataset was created to provide a controlled, reproducible evaluation of character-level reasoning in LLMs. Difficulty tiers allow fine-grained analysis of how word length and letter frequency affect performance.

Evaluation

Scoring is exact match on the integer answer. A fuzzy parser handles written-out number words ("three" → 3). Partial credit is not awarded.

from datasets import load_dataset

ds = load_dataset("floleuerer/strawberry-bench")

# Evaluate a model
correct = sum(1 for ex in ds["test"] if model_predict(ex["question"]) == ex["answer"])
accuracy = correct / len(ds["test"])

License

Creative Commons Attribution 4.0 International (CC BY 4.0)