David Hall: post-training lacks a public data resource like pretraining's web corpus
What happened
On November 3, 2026, Hacker News featured David Hall's analysis of the gap in public data resources for post-training. He says post-training has no public data resource comparable to the internet corpus used for pretraining. Human activity has to be turned into repeatable reinforcement learning (RL) tasks, and each task needs an environment, an intent, tools and a grader. As an example, the original SWE-bench built 2,294 tasks from more than 90,000 PRs, which shows how hard it is to rebuild environments and verify tasks.
Written by AI from the coverage · updated 1 hour ago
Coverage
Follow the reports to see the story from different sides.
- Hacker News front pageThere is no posttraining data commons
David Hall 指出，后训练缺乏类似预训练互联网语料的公共数据资源库，需要主动将人类活动转化为可重复执行的 RL 任务。任务必须具备环境、意图、工具和评分器，原始 SWE-bench 从超过 90,000 个 PR 中最终构建出 2,294 个任务，体现了重建环境与验证任务的难度。
Heat over time
Not enough continuous observations to draw a trend yet.