9). MAmmoTH2 – harvest 10 million naturally existing instruction data from the pre-training web corpus to enhance LLM reasoning; the approach first recalls relevant documents, extracts instruction-response pairs, and then refines the extracted pairs using open-source LLMs;

→ View original post on X — @dair_ai