Open Athena author David Hall analyzes why there is no public pool for post-training data

Open Athena author David Hall points out that pretraining has ready-made public corpora such as Common Crawl, whereas post-training consumes tasks rather than documents, so there is no corresponding public data pool. He uses SWE-bench as an example to illustrate the cost of task creation: among more than 90,000 PRs from 12 Python repositories, only 2,294 became original tasks, and after OpenAI manually reviewed 1,699 of them, only 500 Verified tasks remained. He believes the open source community needs to systematically transcribe human activity into RL tasks; otherwise, the ability to train frontier models will be concentrated in the hands of a few companies.