INTELLECT-MATH: Frontier Mathematical Reasoning through Better Initializations for Reinforcement Learning

INTELLECT-MATH is a 7B parameter model optimized for mathematical reasoning. It was trained in two stages, an SFT stage, in which the model was fine-tuned on verified QwQ outputs, and an RL stage, in which the model was trained using the PRIME-RL recipe. We demonstrate that the quality of our SFT data can impact the performance and training speed of the RL stage: Due to its… See the full description on the dataset page: https://huggingface.co/datasets/PrimeIntellect/INTELLECT-MATH-SFT-Data.

About the Math-Students Dataset

This dataset, originally sourced from the UCI Machine Learning Repository, offers a rich collection of data on student performance in a math program. It provides detailed insights into both the academic achievements and the socio-demographic backgrounds of the students, making it an excellent resource for educational data mining and predictive analytics.

Key Features & Attributes

• Demographics & Background:

  • School: Identifies the student's school (e.g., Gabriel Pereira or Mousinho da Silveira).
  • Sex & Age: Basic demographic information to help explore performance trends among different groups.
  • Address & Family Size: Details about the student’s home environment, including whether they live in an urban or rural area and their family size.

• Parental & Household Information:

  • Parental Cohabitation & Education: Data on whether parents live together and their education levels, which can correlate with student support and academic outcomes.
  • Parental Occupation: Information on the mother’s and father’s jobs, providing further context on socioeconomic factors.

• Educational & Behavioral Variables:

  • Study Time & Failures: Weekly study time and history of past class failures help gauge academic dedication and potential challenges.
  • Support & Extracurricular Activities: Records on whether the student has received extra educational support or participates in extracurricular activities, which can influence overall performance.
  • School-Related Factors: Travel time to school, attendance (absences), and participation in additional paid classes contribute to a holistic view of the educational environment.

• Lifestyle & Social Factors:

  • Internet Access, Free Time & Socializing: Variables like internet availability, free time, and how often students go out with friends help capture lifestyle and behavioral patterns.
  • Health & Well-being: Self-reported health status and alcohol consumption patterns during weekdays and weekends offer insights into personal well-being, which may impact academic performance.

• Academic Performance:

  • Grades: The dataset includes three key assessments—G1 (first period grade), G2 (second period grade), and G3 (final grade). G3, the final grade, serves as the primary target variable for predictive models.

Potential Applications

• Predictive Modeling:
  Researchers and data scientists can build regression models to predict final grades (G3) based on the numerous socio-demographic and educational features.
• Exploratory Data Analysis:
  The dataset is ideal for exploring relationships between family background, lifestyle choices, and academic success. For example, one could analyze how study time or parental education levels correlate with performance.
• Educational Interventions:
  By identifying key factors that contribute to academic outcomes, educators and policymakers can develop targeted interventions to support at-risk students.
• Comparative Studies:
  While this dataset focuses on math scores, its structure is similar to the Portuguese language course dataset. This similarity provides opportunities for cross-domain comparisons in educational research.

Additional Insights

• Data Complexity & Quality:
  Despite its moderate size, the dataset is rich in both categorical and numerical variables. This diversity requires careful preprocessing and feature engineering but also offers the chance to uncover complex interactions between various factors.
• Research Impact:
  The dataset has been widely used in the field of educational data mining. Its comprehensive nature has provided a basis for numerous studies examining the interplay between academic performance and a range of external factors.
• Historical Context:
  Originating from a study presented at the 5th FUBUTEC 2008 conference, the dataset has contributed valuable insights into secondary school performance and continues to serve as a benchmark for educational analytics research.

The preference dataset is derived from the stack exchange dataset which contains questions and answers from the Stack Overflow Data Dump. This contains questions and answers for various topics. For this work, we used only question and answers from math.stackexchange.com sub-folder. The questions are grouped with answers that are assigned a score corresponding to the Anthropic paper: score = log2 (1 + upvotes) rounded to the nearest integer, plus 1 if the answer was accepted by the questioner… See the full description on the dataset page: https://huggingface.co/datasets/prhegde/preference-data-math-stack-exchange.

MATH is a new dataset of 12,500 challenging competition mathematics problems. Each problem in MATH has a full step-by-step solution which can be used to teach models to generate answer derivations and explanations.

Explore Math you can really use - every day through data • Key facts: author, publication date, book publisher, book series, book subjects • Real-time news, visualizations and datasets

This dataset is about book subjects, has 4 rows. and is filtered where the books is Essential math for data science : take control of your data with fundamental linear algebra, probability, and statistics. It features 10 columns including book subject, number of authors, number of books, earliest publication date, and latest publication date. The preview is ordered by number of books (descending).

This report includes results for the New York State Math exams for the years 2006-2012. For the results for the New York State Math exams for the years 2013-2023, please follow this link.

Usable Math is a free & open interactive website where you'll find learning modules designed to develop mathematical problem solving skills among young learners in grades 3 to 7. Visit UsableMath.org.

pe-nlp/Skywork-OR1-RL-Data-Math dataset hosted on Hugging Face and contributed by the HF Datasets community

Third grade English Language Arts (ELA) and Math test results for the 2016-2017 school year by census tract for the state of Michigan. Data Driven Detroit obtained these datasets from MI School Data, for the State of the Detroit Child tool in July 2017. Test results were originally obtained on a school level and aggregated to census tract by Data Driven Detroit. Student data was suppressed when less than five students were tested per school.Click here for metadata (descriptions of the fields).

Explore Math for meds : dosages and solutions through data • Key facts: author, publication date, book publisher, book series, book subjects • Real-time news, visualizations and datasets

Historical price and volatility data for US Dollar in MATH across different time periods.

Guatemala: PISA math scores: The latest value from 2022 is 344.199 index points, unavailable from index points in . In comparison, the world average is 439.569 index points, based on data from 78 countries. Historically, the average for Guatemala from 2022 to 2022 is 344.199 index points. The minimum value, 344.199 index points, was reached in 2022 while the maximum of 344.199 index points was recorded in 2022.

Explore A basic math approach to concepts of chemistry through data • Key facts: author, publication date, book publisher, book series, book subjects • Real-time news, visualizations and datasets

Mathematics scores for Year 9 students.

Historical price and volatility data for Math-e-MATIC in US Dollar across different time periods.

Explore Making maths meaningful through data from visualizations to datasets, all based on diverse sources.

Abstract Issues involving the identification, analysis, and interpretation of mistakes made by mathematics students are not recent, although much still can be investigated on this subject. The aim of the present study was to identify, from the existing literature, relevant variables in the production of errors in mathematics. A systematic review of the literature of the period between 2012 and 2017 was independently performed by two researchers to evaluate the concordance between them. We searched the ERIC, PsycArticles, SciELO and Math Educ Database databases with the descriptors error AND mathematics OR math, error AND procedure AND mathematics OR math, error pattern AND mathematics OR math, analysis of errors AND mathematics OR math, systemic error AND mathematics OR math and their correspondents in Portuguese and Spanish. A total of 415 articles were identified, of which 31 were analyzed, dealing with error production. The variables identified as responsible for producing the most common errors refer to the student's internal causes or unspecified difficulties and errors in the teaching procedures. Responsibility for error is usually attributed to the students and the main trend of the research is only to inform the production of errors, since only a few studies have indicated ways to avoid or deal with errors produced by students in a specific and descriptive way. We emphasize importance and necessity of investigating educational practices to prevent and deal with errors.

Explore How to do maths through data • Key facts: author, publication date, book publisher, book series, book subjects • Real-time news, visualizations and datasets