Datasets:

CJJones
/

Word_letters_Synthetic_Dialog

This dataset contains 262,000 question-answer pairs focused on counting letter occurrences in English words. Each entry features a natural language question asking about the frequency of a specific letter in a given word, paired with a grammatically correct answer. The dataset was curated from a vocabulary extracted from a large-scale hybrid database, with human inspection applied to ensure quality and balance.

Key Statistics:

Total Records: 262,000

Unique Words: ~262,000

Question Variations: 50 paraphrased templates

Answer Variations: 50 paraphrased templates

Average Word Length: 6-8 characters

Output Format: User: [question] Me: [answer]

Dataset Structure Data Fields Each instance in the dataset consists of:

Field Type Description user_query string The natural language question asking about letter frequency assistant_response string The answer providing the letter count with proper grammar word string The target word being analyzed character string The letter being counted count integer Number of occurrences of the character in the word Data Example json { "user_query": "How many e are there in development?", "assistant_response": "There are 3 e in development.", "word": "development", "character": "e", "count": 3 } json { "user_query": "What is the frequency of o in notebook?", "assistant_response": "I'm seeing 3 o in notebook.", "word": "notebook", "character": "o", "count": 3 } json { "user_query": "How many times does a appear in application?", "assistant_response": "I count 2 occurrences of a in application.", "word": "application", "character": "a", "count": 2 } Dataset Creation Data Source The vocabulary was extracted from the full_hybrid table of the Export_Full_Hybrid_Copy.db database, which contains tens of millions of entries. The extraction process:

Batch Processing: Processed the database in 100,000-row batches

Word Extraction: Used regex pattern \b[a-zA-Z]+\b to extract words from question and answer fields

Deduplication: Applied HashSet-based deduplication for O(1) lookup performance

Filtering: Only words between 2-30 characters were retained for quality control

Curation Process The dataset underwent human inspection to ensure:

Grammatical Correctness: All questions and answers use proper English grammar

Logical Consistency: Answers accurately reflect the actual letter counts

Variety Balance: Balanced distribution of:

Word lengths (2-30 characters)

Letter frequencies (0-9 occurrences per word)

Question/answer template diversity

Domain and difficulty levels

Dataset Balance After human inspection, the dataset was trimmed to 262,000 records with:

Templates: 50 question templates × 50 answer templates = 2,500 possible combinations

Word Diversity: Range of common to uncommon English words

Character Coverage: All 26 letters of the English alphabet

Count Distribution:

The dataset uses 50 paraphrased question templates, including:

text

"How many %c are there in %w?"
"What is the count of %c in %w?"
"How many times does %c appear in %w?"
"Can you tell me how many %c are in %w?"
"What's the number of %c in %w?"
"Count the %c in %w for me."
"How many occurrences of %c are in %w?"
"Could you count the %c in %w?"
"What is the frequency of %c in %w?"
"How many %c can be found in %w?" ... (40 more templates) Answer Templates (50 Variations) The dataset uses 50 paraphrased answer templates with proper grammar:

text

"There %s %d %c in %w."
"The count is %d %c in %w."
"I count %d %c in %w."
"There are %d instances of %c in %w."
"The number is %d for %c in %w."
"I see %d %c in %w."
"The total is %d %c in %w."
"There %s exactly %d %c in %w."
"The frequency of %c in %w is %d."
"I find %d %c in %w." ... (40 more templates) Note: The %s placeholder is dynamically replaced with:

"is" when count = 1

"are" when count ≠ 1 (including 0)

Dataset Usage Intended Uses This dataset is designed for:

Natural Language Understanding (NLU): Training models to understand counting and quantitative reasoning questions

Question Answering Systems: Fine-tuning models for arithmetic and counting tasks

Few-Shot Learning: Providing diverse examples for prompt engineering

Grammar and Syntax Training: Learning proper English grammar (is/are agreement)

Educational Applications: Teaching letter recognition and counting skills

Potential Use Cases Chatbot Training: Improving conversational AI's ability to handle quantitative queries

Educational Software: Building literacy and numeracy learning tools

Language Model Fine-tuning: Enhancing models on counting and arithmetic tasks

Data Augmentation: Supplementing existing datasets with diverse phrasing

Limitations and Biases Vocabulary Scope: Limited to words extracted from the source database, may not cover all English words

Word Length Bias: Focus on 2-30 character words, excluding very short or extremely long words

Letter Distribution: Natural language distribution of letters may not be perfectly balanced

Template Diversity: While 50 variations are provided, human-like creativity is not fully replicated

Counting Complexity: Only single-character counting, not multi-character sequences or patterns

Citation If you use this dataset in your research, please cite it as:

bibtex @dataset{letter_count_qa_2024, author = {Letter Count Dataset Team}, title = {Letter Count Q&A: A Dataset for Counting Letter Occurrences in Words}, year = {2024}, publisher = {CJ Jones}, howpublished = {\url{https://huggingface.co/datasets/letter-count-qa}}, description = {262,000 question-answer pairs for letter counting with 50 paraphrased variations each for questions and answers}, }