Share from 0:00
0:00
/

"A Graph-Based Synthetic Data Pipeline for Scaling High-Quality Reasoning Instructions"

The podcast on this paper is generated with Google's Illuminate.

Graph-based Synthetic Data Pipeline (GSDP) turns math knowledge into graphs to generate millions of unique practice problems.

A novel graph-based pipeline that generates high-quality math reasoning data by exploring relationships between knowledge points, achieving 255x data expansion at 1% cost.

-----

https://arxiv.org/abs/2412.08864

πŸ€” Original Problem:

β†’ Current synthetic data approaches for mathematical reasoning suffer from limited scalability, high costs using GPT models, and generate problems too similar to seed data.

-----

πŸ”§ Solution in this Paper:

β†’ The Graph-based Synthetic Data Pipeline (GSDP) extracts knowledge points from seed math problems using specialized models.

β†’ It builds a Knowledge Point Relationships Graph (KPRG) where nodes are knowledge concepts and edges show their connections.

β†’ The system explores both explicit relationships (directly connected points) and implicit relationships (points connected through multiple hops).

β†’ Multiple open-source models jointly evaluate and filter the generated problems and solutions.

-----

πŸ’‘ Key Insights:

β†’ Implicit knowledge relationships enable much higher data expansion than explicit ones

β†’ Using open-source models for synthesis and evaluation drastically reduces costs

β†’ Graph-based approach generates more diverse problems than seed-based methods

-----

πŸ“Š Results:

β†’ Generated 1.91M high-quality math problems from just 7.5K seed problems (255x expansion)

β†’ Achieved synthesis quality comparable to GPT-4 at less than 1% of the cost

β†’ GSDP-7B model reached 37.7% accuracy on MATH and 78.4% on GSM8K benchmarks

Discussion about this video

User's avatar

Ready for more?