Dolphin Math Data Generator: Synthetic Math Problems for Training Language Models in Multi-Step Reasoning
Summary
This project introduces Dolphin Math Data Generator, a tool for creating synthetic math problems and detailed step-by-step solutions across arithmetic, algebra, geometry, and statistics. The generated data is designed to train language models in multi-step mathematical reasoning, supporting both Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) approaches. The project covers 114 math skills and emphasizes generating solutions that mimic human problem-solving processes with a visible scratchpad approach.
Source
Twitter / XDolphin Math Data Generator: Synthetic Math Problems for Training Language Models in Multi-Step Reasoninggithub.comKey quotes
· 3 pulledIt also generates detailed, step-by-step solutions intended to mimic the process a human would follow when solving the problem manually (like a 'visible scratchpad').
The output is designed for training language models to perform multi-step mathematical reasoning.
It works for both SFT and RL. You should generate separate datasets for SFT and RL. SFT teaches it the syntax, RL teaches it to git gud at it.
You might also wanna read
Transformers Can Learn to Predict Permuted Congruential Generator Sequences Through Curriculum Learning and Scaling Laws
This research paper investigates whether Transformer models can learn to predict sequences generated by Permuted Congruential Generators (PC
arXiv.org·2mo ago
Google DeepMind's Aletheia: An Autonomous AI System for Mathematical Research and Proof Generation
Google DeepMind researchers introduce Aletheia, an autonomous mathematics research agent that can generate, verify, and revise mathematical
arxiv.org·4mo ago
Coding Agent Formalization of Numerical Analysis Reveals Flaws in Compilation-Based Quality Metrics
This paper presents research on using coding agents to formalize advanced mathematics textbooks in Lean 4, specifically focusing on Numerica
arxiv.org·18d ago
DeepSeek-Math-V2: Advancing Mathematical Reasoning with Self-Verification Capabilities
DeepSeek-Math-V2 is a new AI model focused on mathematical reasoning that introduces a self-verification approach to overcome limitations of
R-Zero: A Self-Evolving LLM Framework That Generates Its Own Training Data Without Human Input
R-Zero is a fully autonomous framework for training self-evolving Large Language Models (LLMs) that generates its own training data from scr
arxiv.org·9mo agoExploring the Use of Randomly Generated Data for Model Pre-Training
The article explores the use of randomly generated data for pre-training models, supported by theoretical justifications and empirical evide
arxiv.org·1y ago
Comments
Sign in to join the conversation.
No comments yet. Be the first.