The LLM Large Model Training Dataset (Mathematics) is a math instruction tuning dataset containing 200 million problem solutions generated by the GPT-4o model.

These problems are sourced from more than 1,000 math platforms in the United States and other regions, as well as OpenAI training subsets, and are comprehensively solved by allowing major model technologies to use a mix of textual reasoning and code blocks executed by the Python interpreter.

The dataset is split into training and validation subsets which we use in our ablation experiments.

The LLM large model training data package (mathematics class) contains the following fields:

• Questions : Original questions from over 1,000 AI platforms and OpenAI training sets around the world.
• generated_solution : A solution generated using a mix of textual reasoning and code blocks.
• expected_answer : The true answer provided in the original dataset.
• predict_answer : The answer predicted by the Mixtral model in the corresponding solution (from which \boxed{} are extracted ).
• error_message : <not_executed> if the code was not used. Otherwise empty or contains a Python exception from the corresponding code block. The string timeout indicates that the code block took more than 10 seconds to execute. In the current dataset version, we always stop generating after any error or timeout.
• is_correct : Whether our scoring script considers the final answer correct.
• Dataset : neuronicx1000 or OpenAI-math.
• generation_type : without_reference_solution or masked_reference_solution .
  
  
  LLM large model training data package (mathematics 500G) usage process

1. Purchase & Download

   • Choose to purchase the LLM large model training data package (mathematics 500G) on the platform.
   • Once payment is complete, you will be notified of the download link or data delivery method.
   • Download the data package to the local storage device.

2. Unzip and organize

   • Once the download is complete, extract the data package, which is usually compressed in ZIP or RAR format.
   • Data files will be classified and organized according to language, academic level (such as middle school, university) and specific fields (such as algebra, geometry, statistics, etc.) for easy search and use.

3. Data preprocessing

   • Format the data according to project requirements and adapt it to your AI model training framework (such as PyTorch, TensorFlow, etc.).
   • Check for noise or non-compliant content in the data to ensure the accuracy of training.

4. Import model training environment

   • Import data into your model training environment .
   • Make sure the data loading meets the input requirements of the model, such as input data format, batch size, etc.

5. Model Training

   • Use this data package for model training. This data package is particularly suitable for multi-language mathematical model training, covering academic mathematics content from middle school to university.
   • Combined with the mathematical knowledge in the data, the model can be applied to multiple fields such as natural language processing, intelligent answering , and problem-solving systems.

6. Optimization and debugging

   • During the training process, adjust the model parameters, optimizer, learning rate, etc. according to the preliminary results to improve the accuracy and performance of the model.
   • Compare the impact of data from different academic fields on the model results to ensure comprehensive coverage of required knowledge points.

7. Output and Application

   • After training, the model will be used in application scenarios, such as solving math problems and intelligent education platforms .
   • The multi-language, multi-level data in the data package supports a wide range of application scenarios, especially AI projects involving the global mathematics field.

With this data package, you will easily obtain high-quality mathematical data in multiple languages and academic levels to empower your AI models.
Release date: September 9, 2024