Meta Introduces Autodata, an Agentic Data Scientist for High-Quality Synthetic Data

Meta has introduced Autodata, a general method where AI agents act as data scientists to create high-quality training and evaluation data. This system can meta-optimize itself to become an even stronger data scientist. Autodata generalizes existing synthetic data creation methods, with a practical implementation called Agentic Self-Instruct, which has shown improved results in tasks like computer science research and legal reasoning. This approach enables leveraging increased inference compute to generate more challenging and higher-quality datasets.