Machine-generated code detection on SemEval Task 13 Subtask A 2026 official (test)
78.9Macro F1SALSA Tuned (Qwen2.5-72B-Instruct)
Evaluation Results
| Method | Links | |
|---|---|---|
| SALSA Tuned (Qwen2.5-72B-Instruct)Tuning strategy=Fine-tuned, Model backbone=Qwen2.5-72B-Instruct2026.06 | 78.9 | |
| SALSA Tuned (Qwen3-32B)Tuning strategy=Fine-tuned, Model backbone=Qwen3-32B2026.06 | 76 | |
| SALSA Tuned (Qwen3-32B official)Tuning strategy=Fine-tuned, Model backbone=Qwen3-32B, Submission status=Official Kaggle submission2026.06 | 73 | |
| gemma-3-27b-pt_th1classification threshold=1.0, parameter size=27B2026.04 | 69.753 | |
| codegemma-1.1-7b-it_th1classification threshold=1.0, parameter size=7B2026.04 | 64.625 | |
| codegemma-7b_th1classification threshold=1.0, parameter size=7B2026.04 | 64.435 | |
| Qwen2.5-Coder-14B_th1classification threshold=1.0, parameter size=14B2026.04 | 63.818 | |
| codegemma-7bclassification threshold=0.5, parameter size=7B2026.04 | 59.628 | |
| SALSA Zero-shot (Qwen3-Next-80B-A3B-Instruct)Tuning strategy=Zero-shot - without tuning, Model backbone=Qwen3-Next-80B-A3B-Instruct2026.06 | 59.1 | |
| Qwen2.5-Coder-7Bclassification threshold=0.5, parameter size=7B2026.04 | 57.799 | |
| CodeLlama-70b-hf_th1classification threshold=1.0, parameter size=70B2026.04 | 56.363 | |
| CodeLlama-70b-hfclassification threshold=0.5, parameter size=70B2026.04 | 54.776 | |
| Qwen2.5-Coder-14Bclassification threshold=0.5, parameter size=14B2026.04 | 54.498 | |
| codegemma-1.1-7b-itclassification threshold=0.5, parameter size=7B2026.04 | 54.183 | |
| Qwen2.5-Coder-32Bclassification threshold=0.5, parameter size=32B2026.04 | 51.694 | |
| SALSA Zero-shot (Qwen3-32B)Tuning strategy=Zero-shot - without tuning, Model backbone=Qwen3-32B2026.06 | 50.8 | |
| random baseline2026.04 | 50 | |
| Qwen2.5-Coder-0.5Bclassification threshold=0.5, parameter size=0.5B2026.04 | 49.474 | |
| gemma-3-27b-ptclassification threshold=0.5, parameter size=27B2026.04 | 48.006 | |
| SALSA Tuned (Qwen3-8B)Tuning strategy=Fine-tuned, Model backbone=Qwen3-8B2026.06 | 45 | |
| SALSA Zero-shot (Qwen2.5-72B-Instruct)Tuning strategy=Zero-shot - without tuning, Model backbone=Qwen2.5-72B-Instruct2026.06 | 44.9 | |
| SALSA Zero-shot (Qwen3-8B)Tuning strategy=Zero-shot - without tuning, Model backbone=Qwen3-8B2026.06 | 43.6 | |
| CodeBERT baseline2026.04 | 30.53 | |
| CodeBERTCategory=Baseline2026.06 | 30.5 |