Developer releases free dataset of recorded AI coding sessions, including failures
A developer is offering 66 recorded coding sessions, comprising 2,541 real tool calls across five open-source repositories, under an MIT license at no cost. The recordings capture full work sessions, including failed test runs and their fixes, rather than just final code diffs. The author argues that AI companies have profited from developers' code and that this data should be shared freely instead of sold. They report that training a small Qwen model on the data raised correct tool calls on their own test from 43% to 80%, while other benchmark scores worsened, and they have published these results openly. The dataset includes a verification script so users can confirm the files match what was published.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in