CodeMidas
A method for automatically converting existing code repositories into executable programming training tasks. By reverse-constructing function implementations, tests, and verifiers, CodeMidas provides a scalable data source for reinforcement learning training of programming agents.
Timeline (last 90 days)
CodeMidas proposes to directly transform the behavior of existing code repositories into executable training tasks.
The name CodeMidas is derived from the Midas touch (turning things to gold).
CodeMidas's approach is analogous to masking a function's implementation while retaining its specification, using the function signature and existing tests as task input to require the model to re-implement it.
CodeMidas pursues automated batch processing of any repository that meets the criteria, transforming data production from a bottleneck into a pipeline.
Current public information mainly describes the overall approach of the CodeMidas method, lacking specific implementation details, conversion quality, and real RL training effect data.
All Facts (4)
CodeMidas proposes to directly transform the behavior of existing code repositories into executable training tasks.
50%UnverifiedThe name CodeMidas is derived from the Midas touch (turning things to gold).
50%UnverifiedCodeMidas's approach is analogous to masking a function's implementation while retaining its specification, using the function signature and existing tests as task input to require the model to re-implement it.
50%UnverifiedCodeMidas pursues automated batch processing of any repository that meets the criteria, transforming data production from a bottleneck into a pipeline.
50%