There is no posttraining data commons
Summary
This explainer argues that there is no posttraining data commons, contrasting the abundance of data used for pretraining with the scarcity of data and well-defined tasks for posttraining. It discusses the challenges of converting human activities into durable RL tasks, the difficulties observed in datasets like SWE-bench, and the call for deliberate creation of a posttraining commons to democratize AI development and governance.