SemEval-2026 Task 13: Fine-tuned CodeBERT with Stratified Balancing, Dynamic Threshold Optimization, and Logit Bias Correction for Robust Multi-Language AI Code Detection
Year: 2026
Type: conference-paper
Abstract: TeamUdayThalavesh participated in SemEval-2026 Task 13 Subtask A, addressing binary classification between human-written and AI-generated code across multiple programming languages, generators, and application scenarios using the AICD-Bench dataset.Our system finetunes microsoft/codebert-basepretrained on CodeSearchNet-on a 200K balanced training subset via stratified sampling.TPU-accelerated training with per-epoch F1-macro threshold optimization and label flip bias correction addresses Code-BERT's overconfidence, achieving validation F1-macro=0.874 (threshold=0.52,epoch 3).Our single-model submission secured private leaderboard F1-macro = 0.53 (rank ∼40).We further conduct a detailed post-hoc error analysis to investigate the substantial performance gap between validation and test distributions, identifying distribution shift across programming languages and generator families as the primary source of degradation.
Institution: Sri Sivasubramaniya Nadar College of Engineering
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166806656