How codebase-specific RL fine-tuning enables smaller, specialized AI models to significantly outperform general-purpose models at generating impactful unit tests.

Our thoughts on SWE-Bench after reading a recent paper that found issues with SWE-Bench resulting in SWE-Agent + GPT 4's performance dropping 3x.

The inner workings of our repo-specific finetuning pipeline that yields 50% more accurate code completions.

Our thesis on why finetuning with internal engineering data sources is crucial for complex, enterprise-grade codebases with unique patterns & frameworks.