KwaiKAT Releases KAT-Coder-V2.5 With Verifiable Repository Training

The KwaiKAT Team at Kuaishou has introduced KAT-Coder-V2.5, an advanced coding model designed to operate within real, executable repository environments rather than simply generating single-turn code. This innovative approach aims to improve the accuracy and reliability of generated code by grounding it in practical, testable contexts. The served model is accessible through StreamLake, while an open-weight variant, KAT-Coder-V2.5-Dev, has been made available on Hugging Face under the permissive Apache-2.0 license.

The research behind KAT-Coder-V2.5 frames a verifiable task as a triplet comprising a precise task description, an executable repository environment, and a comprehensive set of validation tests. A code patch is deemed correct only if it successfully passes all these tests. To achieve this, tasks are sourced from real-world pull requests and commits, following the methodology established by the SWE-bench project. The merged code change serves as the "golden patch," and any accompanying test changes provide the "test patch." Raw issue text is not used as a specification; instead, descriptions are regenerated into three distinct parts: a problem statement directly linked to the golden patch, requirements derived from the test patch, and interface constraints inferred from both. A subsequent clarity check filters out any information that is ambiguous, incomplete, underspecified, or internally inconsistent.

The AutoBuilder system handles the crucial environment aspect of this process. A build agent meticulously analyzes the repository to generate a configuration script that manages dependency installation and test execution from a clean checkout. Subsequently, a verification agent executes this script within an isolated sandbox environment. The acceptance criteria for a valid environment are particularly noteworthy. Verification does not rely on simple exit codes or log pattern matching. Instead, it parses structured output from test frameworks and accepts an environment only when more than 90% of the expected tests are collected and their pass/fail outcomes can be reproduced across multiple runs. Any identified failures are fed back as structured information to facilitate iterative repair of the code or environment.

By combining a preconfigured base environment, templates for build systems, and a retrievable library of distilled build recipes, the system significantly improved its construction success rate from an initial 16.5% to 57.2%. This robust methodology has resulted in the creation of over 100,000 verifiable environments, encompassing code written in 12 different programming languages. This extensive dataset, derived from real Git history and commit data, provides a rich and practical foundation for training and evaluating coding models like KAT-Coder-V2.5, pushing the boundaries of agentic AI in software development.