This is a reformatted version of OpenAI's GSM8K (Grade School Math 8K) dataset, converted into an instruction-following format suitable for training and evaluating large language models on mathematical reasoning tasks.

GSM8K is a dataset of 8.5K high-quality, linguistically diverse grade school math word problems created by human problem writers. The problems require multi-step reasoning and basic arithmetic operations (addition, subtraction, multiplication, division) to solve. A bright middle school student should be able to solve every problem in this dataset.

Key Features:

📊 7,473 training examples + 1,319 test examples
🎯 Instruction-following format ready for LLM fine-tuning
📝 Step-by-step solutions with reasoning chains
🔢 Clean extraction of final numeric answers
🎓 Includes both standard and Socratic (guided questioning) formats

Original Source

This dataset is derived from:

Original Repository: openai/grade-school-math
Paper: Training Verifiers to Solve Math Word Problems
Authors: Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, et al. (OpenAI)
License: MIT License

Supported Tasks

Instruction Following: Training models to follow mathematical problem-solving instructions
Chain-of-Thought Reasoning: Learning to break down complex problems into steps
Mathematical Reasoning: Multi-step arithmetic problem solving
Question Answering: Extracting and computing numerical answers
Educational AI: Building tutoring systems and educational assistants

Languages

The dataset is in English (en).

Dataset Structure

Data Format

Each example in the instruction format contains:

{
  "id": "gsm8k_train_0001",
  "source": "GSM8K",
  "split": "train",
  "instruction": "Solve the following math word problem. Show your work step by step.",
  "input": "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?",
  "output": "Natalia sold 48/2 = 24 clips in May.\nNatalia sold 48+24 = 72 clips altogether in April and May.",
  "final_answer": "72"
}

Data Fields

id: Unique identifier for the example (e.g., "gsm8k_train_0001")
source: Dataset source ("GSM8K" or "GSM8K-Socratic")
split: Data split ("train" or "test")
instruction: The instruction prompt for the model
input: The math word problem to solve
output: Step-by-step solution showing reasoning
final_answer: The final numeric answer (extracted from the solution)

Data Splits

Split	Standard	Socratic	Total
Train	7,473	7,473	14,946
Test	1,319	1,319	2,638
Total	8,792	8,792	17,584

Files

Standard Format

gsm8k_train.jsonl - Training set with standard solutions
gsm8k_test.jsonl - Test set with standard solutions

Socratic Format

gsm8k_train_socratic.jsonl - Training set with Socratic subquestions
gsm8k_test_socratic.jsonl - Test set with Socratic subquestions

The Socratic format includes automatically generated subquestions before each reasoning step, designed to guide the problem-solving process more explicitly.

Dataset Creation

Source Data Curation

The original GSM8K dataset was created by OpenAI with human problem writers who crafted linguistically diverse word problems that:

Take between 2 and 8 steps to solve
Require only basic arithmetic operations
Are appropriate for grade school level (ages 6-14)
Test multi-step reasoning ability

Conversion Process

This instruction-format version was created by:

Cleaning: Removed calculator annotations (e.g., <<48/2=24>>) from solutions
Parsing: Extracted step-by-step reasoning and final numeric answers
Reformatting: Structured data into instruction-input-output format
Standardization: Added metadata (source, split, unique IDs)

The conversion preserves all original problems and solutions while making the format more suitable for modern instruction-tuned language models.

Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("notefill/gsm8k-instruction")

# Load specific splits
train_data = load_dataset("notefill/gsm8k-instruction", data_files="gsm8k_train.jsonl")
test_data = load_dataset("notefill/gsm8k-instruction", data_files="gsm8k_test.jsonl")

# Load Socratic version
train_socratic = load_dataset("notefill/gsm8k-instruction", data_files="gsm8k_train_socratic.jsonl")

Fine-tuning Example

from datasets import load_dataset

# Load training data
dataset = load_dataset("notefill/gsm8k-instruction", data_files="gsm8k_train.jsonl")

# Format for instruction tuning
def format_prompt(example):
    return {
        "prompt": f"{example['instruction']}\n\n{example['input']}",
        "completion": f"{example['output']}\n\nFinal Answer: {example['final_answer']}"
    }

formatted_dataset = dataset.map(format_prompt)

Evaluation Example

import json

def extract_final_answer(text):
    """Extract numeric answer from model output"""
    # Your extraction logic here
    pass

def evaluate_accuracy(predictions, ground_truth):
    """Calculate accuracy of predictions"""
    correct = 0
    for pred, truth in zip(predictions, ground_truth):
        if extract_final_answer(pred) == truth['final_answer']:
            correct += 1
    return correct / len(predictions)

Problem Characteristics

Difficulty Distribution

Steps Required: 2-8 steps per problem
Operations: Addition, subtraction, multiplication, division
Context: Real-world scenarios (shopping, cooking, travel, etc.)
Complexity: Progressive difficulty suitable for grade school students

Example Problems

Simple (2-3 steps):

"Weng earns $12 an hour for babysitting. Yesterday, she just did 50 minutes of babysitting. How much did she earn?"

Medium (4-5 steps):

"Betty is saving money for a new wallet which costs $100. Betty has only half of the money she needs. Her parents decided to give her $15 for that purpose, and her grandparents twice as much as her parents. How much more money does Betty need to buy the wallet?"

Complex (6-8 steps):

"Mark has a garden with flowers. He planted plants of three different colors in it. Ten of them are yellow, and there are 80% more of those in purple. There are only 25% as many green flowers as there are yellow and purple flowers. How many flowers does Mark have in his garden?"

Considerations for Using the Data

Recommended Uses

✅ Training instruction-following language models
✅ Evaluating mathematical reasoning capabilities
✅ Developing chain-of-thought prompting strategies
✅ Building educational AI tutoring systems
✅ Research in multi-step reasoning
✅ Benchmarking model arithmetic abilities

Limitations

⚠️ Calculation Errors: Models may struggle with arithmetic despite correct reasoning
⚠️ Problem Complexity: Limited to grade school difficulty
⚠️ Domain Coverage: Focuses on word problems, not pure mathematics
⚠️ Language: English only
⚠️ Cultural Context: May reflect specific cultural references

Ethical Considerations

The dataset is designed for research and educational purposes
Solutions demonstrate reasoning processes that can be learned and replicated
Care should be taken when deploying models trained on this data in educational settings
Human oversight is recommended for student-facing applications

Benchmark Performance

From the original GSM8K paper, model performance on the test set:

Model	Accuracy
GPT-3 175B	34.2%
GPT-3 175B (Verification)	55.0%
GPT-3 6B	19.7%
GPT-3 6B (Verification)	33.4%

(Note: These are results from the original 2021 paper. Newer models may perform differently.)

Citation

Original GSM8K Dataset

@article{cobbe2021gsm8k,
  title={Training Verifiers to Solve Math Word Problems},
  author={Cobbe, Karl and Kosaraju, Vineet and Bavarian, Mohammad and Chen, Mark and Jun, Heewoo and Kaiser, Lukasz and Plappert, Matthias and Tworek, Jerry and Hilton, Jacob and Nakano, Reiichiro and Hesse, Christopher and Schulman, John},
  journal={arXiv preprint arXiv:2110.14168},
  year={2021}
}

This Instruction Format Version

@dataset{gsm8k_instruction2025,
  title={GSM8K: Grade School Math 8K (Instruction Format)},
  author={Kuyeso Rogers and Adiza Alhassan and Notefill},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/datasets/notefill/gsm8k-instruction}},
  note={Instruction-following format conversion of OpenAI's GSM8K dataset}
}

Licensing Information

This dataset retains the MIT License from the original GSM8K dataset.

MIT License

Copyright (c) 2021 OpenAI

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.

Additional Resources

Original Repository: github.com/openai/grade-school-math
Original Paper: arxiv.org/abs/2110.14168
OpenAI Blog Post: openai.com/blog/grade-school-math

Acknowledgments

We gratefully acknowledge:

OpenAI for creating and releasing the original GSM8K dataset
The original authors: Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman
The human problem writers who created these high-quality educational problems

Contact

For questions or feedback about this instruction-format version, please open an issue on the dataset repository.

For questions about the original GSM8K dataset, please refer to the original repository.

Downloads last month: 13