Autodata

Autodata is a research method from Meta's RAM (Reasoning, Alignment, and Memory) team in which an LLM agent acts as a data scientist. It creates training or evaluation data grounded in provided documents, analyzes the results at example and dataset level, and updates its data-generation recipe over repeated iterations until a stopping criterion is met. It generalizes earlier synthetic-data approaches such as Self-Instruct, Grounded Self-Instruct, CoT Self-Instruct and Self-Challenging, which control quality mainly through filtering or refinement.

The authors also describe meta-optimizing the data scientist agent itself (an outer loop), for example by optimizing the agent harness with autoresearch or meta-harness style methods, using the same data-quality criteria as the inner loop. The initial instantiation, Agentic Self-Instruct, uses a main agent with four subagents: a challenger that creates examples, a weak solver, a strong solver, and a verifier/judge. Meta reports gains on scientific reasoning problems over classical synthetic data creation, with a further uplift from meta-optimization. A YouTube summary also mentions a Qwen 3.5 4B model trained on this data outperforming a standard method.

Autodata is a research method published as a blog post and technical report, not a packaged product. The provided material gives no code repository, model weights or hosted service, so access is limited to reading the write-up.