CoolFace
Datasetpublic

vibingshu/crisp_llm_prompt_responses

Crisp LLM Prompt Responses Dataset Overview This dataset contains 100,000 carefully crafted prompt-response pairs which can be used to train Open Source Large Language Models (LLMs) to follow precise formatting instructions and provide crisp, accurate responses. The dataset focuses on instruction-following behavior where models must adhere to specific output constraints. Dataset Characteristics Total Size 100,000 instruction-response… See the full description on the dataset page: https://huggingface.co/datasets/vibingshu/crisp_llm_prompt_responses.

sourceHugging Faceupdated 1y agoView on Hugging Face
0likes19downloads
Dataset Card

Crisp LLM Prompt Responses Dataset

Overview

This dataset contains 100,000 carefully crafted prompt-response pairs which can be used to train Open Source Large Language Models (LLMs) to follow precise formatting instructions and provide crisp, accurate responses. The dataset focuses on instruction-following behavior where models must adhere to specific output constraints.

Dataset Characteristics

Total Size

  • —100,000 instruction-response pairs
  • —Balanced distribution across multiple format types
  • —Diverse topics covering science, mathematics, geography, technology, and general knowledge

Format Distribution

The dataset includes the following format categories:

Format TypeCountPercentageDescription
Yes/No20,00020.0%Binary yes/no questions requiring precise answers
Single Word15,00015.0%Questions requiring one-word responses
Number Only11,17911.2%Mathematical calculations with numeric-only outputs
Binary8,3208.3%Questions requiring 1/0 responses
List Format6,7726.8%Requests for bullet points, comma-separated, or numbered lists
Two Sentences6,3086.3%Explanations limited to exactly two sentences
JSON Format5,0005.0%Structured data responses in JSON format
One Sentence4,7914.8%Explanations limited to exactly one sentence
Word Limited2,0002.0%Responses with specific word count constraints
Code Format1,5151.5%Programming code snippets without explanations

Key Features

1. Format Precision Training

  • —Strict adherence to output format requirements
  • —Clear constraint specification in prompts
  • —Consistent response formatting

2. Topic Diversity

  • —Geography: Capital cities, countries, landmarks
  • —Mathematics: Arithmetic operations, percentages, sequences
  • —Science: Chemical elements, physics, biology facts
  • —Technology: Programming concepts, AI terminology
  • —History: Historical events, dates, figures
  • —Language: Grammar, vocabulary, translations

3. Constraint Variety

  • —Length Constraints: Word limits, sentence counts
  • —Format Constraints: JSON, lists, code blocks
  • —Response Type: Binary, numeric, categorical
  • —Style Constraints: Professional, concise, structured

Example Categories

Single Word Responses

Prompt: "What is the capital of France? One word only."
Response: "Paris"

Binary Responses

Prompt: "Is the Earth round? Answer 1 for yes, 0 for no."
Response: "1"

Mathematical Operations

Prompt: "What is 15 * 8? Number only."
Response: "120"

List Formatting

Prompt: "List the primary colors. Use bullet points."
Response: "- Red
- Blue
- Yellow"

JSON Responses

Prompt: "Create a person object. JSON format only."
Response: "{
  "name": "John Smith",
  "age": 30,
  "city": "New York"
}"

Training Benefits

For LLM Fine-tuning

  1. 1.Format Compliance: Teaches models to follow exact formatting instructions
  2. 2.Constraint Adherence: Improves ability to respect output limitations
  3. 3.Precision Training: Reduces verbose or off-topic responses
  4. 4.Instruction Following: Enhances general instruction-following capabilities

Use Cases

  • —Training smaller, focused models for specific tasks
  • —Fine-tuning existing models for better instruction adherence
  • —Creating specialized assistants with precise output requirements
  • —Improving model reliability in production environments

Quality Assurance

Data Validation

  • —All mathematical calculations verified for accuracy
  • —Factual information cross-checked for correctness
  • —Format consistency maintained across examples
  • —Balanced representation of different constraint types

Response Accuracy

  • —Geographical facts verified (capitals, countries)
  • —Scientific information validated (chemical symbols, physics)
  • —Mathematical operations double-checked
  • —Programming code syntax verified

Usage Recommendations

For Training

  • —Use full 100K dataset for comprehensive fine-tuning
  • —Split 80/10/10 for train/validation/test sets
  • —Shuffle data to ensure format variety in each batch
  • —Monitor format compliance during training

For Evaluation

  • —Test on held-out examples from each format category
  • —Measure exact format match accuracy
  • —Evaluate content accuracy alongside format compliance
  • —Use sample dataset for quick validation

Technical Specifications

Data Format

  • —JSON format with "prompt" and "response" fields
  • —UTF-8 encoding for international character support
  • —Consistent field naming across all examples

Size Statistics

  • —Average prompt length: ~40 characters
  • —Average response length: ~25 characters
  • —Total file size: ~15MB (formatted JSON)
  • —Memory efficient for most training scenarios

License and Attribution

This dataset is designed for research and educational purposes in improving LLM instruction-following capabilities. When using this dataset, please cite appropriately and ensure responsible AI development practices.