“An ideal dataset for advancing AI capabilities would involve multi-turn user interactions within a million-line codebase, documenting a multi-week research task that requires multiple pull requests.”