Fine-Tuning Pre-Trained Code Models for AI-Generated Code Detection

TL;DR

This paper presents a system for detecting AI-generated code using fine-tuned pre-trained code models, achieving competitive results in a shared task.

Contribution

It introduces strategies for fine-tuning multiple pre-trained code models for AI-generated code detection tasks.

Findings

Achieved macro-F1 score of 0.737 on Subtask-A

Achieved macro-F1 score of 0.422 on Subtask-B

Placed 6th and 7th in respective subtasks among many teams

Abstract

This paper describes the system submitted by team \textbf{Archaeology} to SemEval-2026 Task~13 on AI-generated code detection. The shared task consists of three subtasks; we participate in Subtask-A (binary classification: human-written vs.\ AI-generated code) and Subtask-B (11-class attribution of the generating model). Starting from a TF-IDF and Logistic Regression baseline, we fine-tune four pre-trained code models (CodeBERT, GraphCodeBERT, UniXcoder, and CodeT5+) with separate strategies for each subtask. For Subtask-A, we use leave-one-language-out cross-validation, code augmentation, chunked inference with trimmed-mean aggregation, and threshold calibration on a difficult dataset. For Subtask-B, we use sandwich token packing, class-balanced loss, and multi-seed ensembling with test-time augmentation. Our best submissions obtain macro-F1 scores of 0.737 on Subtask-A (6th/81 teams) and 0.422 on Subtask-B (7th/34 teams).