R2C2-Coder

• 📙Paper: R2C2-Coder Enhancing and Benchmarking Real-world Repository-level Code Completion Abilities of Code Large Language Models
• 📚Publisher: arxiv
• 🏠Author Affiliation: Alibaba Group
• Contributions:
  • We investigate the limitations of existing repository-level code completion (e.g., lack of sufficient context and benchmark with limited scenarios) and propose the R2C2-Coder including R2C2-Enhance and R2C2-Bench to enhance and benchmark real-world repositorylevel code completion abilities of Code LLMs.
  • For R2C2-Enhance, we propose to construct the candidate retrieval pool with abstract and snippet contexts and generate the completion prompt using context retrieval and prompt assemble. Based on R2C2-Enhance, we build a new repository-level code completionbenchmark called R2C2-Bench with training, validation, and testing splits, where a context perturbation strategy is used to simulate the real-world completion scenes better.
  • Comprehensive experimental results on multiple benchmark datasets demonstrate the effectiveness and efficiency of our R2C2-Coder.