Smaller but Better: Self-Paced Knowledge Distillation for Lightweight yet Effective LCMs

Large code models (LCMs) have remarkably advanced the field of code generation. Despite their impressive capabilities, they still face practical deployment issues, such as high inference costs, limited accessibility of proprietary LCMs, and adaptability issues of ultra-large LCMs. These issues highlight the critical need for more accessible, lightweight yet effective LCMs. Knowledge distillation (KD) offers a promising solution, which transfers the programming capabilities of larger, advanced LCMs (Teacher) to smaller, less powerful LCMs (Student). However, existing KD methods for code intelligence often lack consideration of fault domain knowledge and rely on static seed knowledge, leading to degraded programming capabilities of student models.

In this paper, we propose a novel Self-Paced knOwledge DistillAtion framework, named SODA, aiming at developing lightweight yet effective student LCMs via adaptively transferring the programming capabilities from advanced teacher LCMs. SODA consists of three stages in one cycle: (1) Correct-and-Fault Knowledge Delivery stage aims at improving the student model's capability to recognize errors while ensuring its basic programming skill during the knowledge transferring, which involves correctness-aware supervised learning and fault-aware contrastive learning methods. (2) Multi-View Feedback stage aims at measuring the quality of results generated by the student model from two views, including model-based and static tool-based measurement, for identifying the difficult questions; (3) Feedback-based Knowledge Update stage aims at updating the student model adaptively by generating new questions at different difficulty levels, in which the difficulty levels are categorized based on the feedback in the second stage. By performing the distillation process iteratively, the student model is continuously refined through learning more advanced programming skills from the teacher model. We compare SODA with four state-of-the-art KD approaches on three widely-used code generation benchmarks with different programming languages. Experimental results show that SODA improves the student model by 65.96% in terms of average Pass@1, outperforming the best baseline PERsD by 29.85%. Based on the proposed SODA framework, we develop SodaCoder, a series of lightweight yet effective LCMs with ∼7B parameters, which outperform 15 LCMs with less than or equal to 16B parameters. Notably, SodaCoder-DS-6.7B, built on DeepseekCoder-6.7B, even surpasses the prominent ChatGPT on average Pass@1 across seven programming languages (66.4 vs. 61.3).