Training code generation models to debug their own outputs

Using large language models to generate training data and updating models through both fine tuning and reinforcement learning improves the success rate of code generation by 39%.Read More