xufana/RedPajama-INCITE-Instruct-3B-Addition

Dataset Summary

The Arithmetic Operations Dataset is a synteticly generated collection of mathematical arithmetic operations for practice and evaluation purposes. It contains a total of 624,800 arithmetic operations, consisting of 568,000 addition operations and 56,800 subtraction operations. The dataset is designed to provide a range of arithmetic problems to train and evaluate language models for solving simple arithmetic (mostly addition, the others TBA) problems.

Dataset Structure

The dataset is organized into two main categories: addition and subtraction. Each category contains a set of arithmetic operations in separate files (addition.json) and (subtraction.json), and the file (dataset.json) provides combined data from both.

Data Instances

{
    "instruction": "What is the answer to 373486002216116154 + 339369?",
    "input": "373486002216116154 + 339369",
    "output": "373486002216116154 + 339369 = 373486002216455523",
    "answer": "373486002216455523"
},
{
    "instruction": "9916607491627649 minus 581954",
    "input": "9916607491627649 - 581954",
    "output": "9916607491627649 - 581954 = 9916607491045695",
    "answer": "9916607491045695"
},

Data Fields

The files share the same structure and have 4 fields:

• instruction: Human instructions are generated by inserting arithmetic expressions into randomly selected templates and incorporating natural language variations. These instructions are intended to serve as prompts for instruction-finetuning, providing input for training the model.
• input: A randomly generated arithmetic expression, that can serve as a substitute for the 'instruction' component during training, allowing a specific focus on arithmetic operations while minimizing the impact of natural language.
• output: the target output for the model to learn.
• answer: direct numerical answer to the arithmetic task. It can be used to test learnability of various sub-tasks.

Contact

For any questions or inquiries regarding this dataset, please contact xufana@yandex.ru.