CodeDet-NITS at SemEval-2026 Task 13: AI Code Authorship Detection Beyond Truncation

Paper · 2026 · Lekkala Sai Teja, Annepaka Yadagiri, Kshitij Patiyal et al.

Summary

The paper tackles fine-grained attribution of source code for SemEval-2026 Task 13 Subtask B. Each snippet must be assigned to one of eleven classes: human-written or one of ten LLM families. Systems are scored by macro F1 because the official data (500,000 training and 100,000 validation samples) is dominated by human code. The authors fine-tune instruction-tuned Qwen2.5-Coder models (3B and 7B) as sequence classifiers using QLoRA. The base weights stay frozen in 4-bit form, and only low-rank adapters plus a classification head are trained. A short language tag is prepended to each input. Their main contribution is handling long snippets without truncation:

Training: each snippet is split into overlapping fixed-length windows that all inherit the snippet's label.
Inference: the windows are created dynamically at prediction time, and the logits from all of them are averaged to produce one prediction.

The system placed 8th on the official test leaderboard. Ablations show that covering the whole snippet and using soft aggregation account for most of the gains. This suggests that authorship signals can appear anywhere in the code, not just at its start.

Key points

The team (CodeDet-NITS, NIT Silchar) ranked 8th on Subtask B with a test macro F1 of 0.41570. The top system (TeleAI) scored 0.508, and the CodeBERT-Baseline scored 0.228.
Training data is heavily imbalanced: about 442.1k of the training samples are human-written, versus roughly 2k–10.8k per LLM family.
Windows are 512 tokens long with a stride of 128, so adjacent windows overlap by 384 tokens. QLoRA uses NF4 quantization and paged AdamW, implemented with Transformers, Accelerate and bitsandbytes.
On validation, the windowed Qwen2.5-Coder-7B QLoRA model performed best (64.61 macro F1, 95.41% accuracy), ahead of the 3B variant (60.32).
Ablations on validation macro F1:

Variant Macro F1

Truncation to the first window 54.87

No language prefix 56.14

Stride 256 57.29

Majority-vote pooling instead of logit averaging 58.03

Variant	Macro F1
Truncation to the first window	54.87
No language prefix	56.14
Stride 256	57.29
Majority-vote pooling instead of logit averaging	58.03

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Institutions

🎓National Institute Of Technology SilcharSilchar ·

India6 works · 72 citations

Fields, topics and concepts

📚 Shelves 5Computer Science · 16,667 works Artificial Intelligence · 16,194 works Arts and Humanities · 16,039 works Language and Linguistics · 15,891 works Information Systems · 1,437 works

Cite

Cite as Lekkala Sai Teja, Annepaka Yadagiri, Kshitij Patiyal, Sangam Sai Anish and Partha Pakray. 2026. CodeDet-NITS at SemEval-2026 Task 13: AI Code Authorship Detection Beyond Truncation. Proceedings of the 20th International Workshop on Semantic Evaluation (2026).

The paper itself, not this page. BibTeX from the library's record of the work. DOI 10.18653/v1/2026.semeval-1.397

@inproceedings{teja2026codedet,
  title = {{CodeDet-NITS at SemEval-2026 Task 13: AI Code Authorship Detection Beyond Truncation}},
  author = {Lekkala Sai Teja, Annepaka Yadagiri, Kshitij Patiyal, Sangam Sai Anish and Partha Pakray},
  booktitle = {Proceedings of the 20th International Workshop on Semantic Evaluation (2026)},
  year = {2026},
  doi = {10.18653/v1/2026.semeval-1.397},
  url = {https://doi.org/10.18653/v1/2026.semeval-1.397}
}

References

6 references, 6 of them with a card of ours. ↗ leads to the work.

\#\# Conclusion We presented our submission to SemEval 2026 Task 13 Subtask B, addressing multi class code authorship detection across human code and ten LLM families. Our system fine tunes Qwen2.5 Coder models using QLoRA and relies on a token complete long context strategy that replaces truncation T…
Zhangyin Feng, Daya Guo, Duyu Tang, Nan Duan, Xiaocheng Feng, Ming Gong, Linjun Shou, Bing Qin, Ting Liu, Daxin Jiang, and 1 others. 2020. Codebert: A pre-trained model for programming and natural languages. In Findings of the association for computational linguistics: EMNLP 2020, pages 1536–1547.
Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. Iclr, 1(2):3.
Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Keming Lu, and 1 others. 2024. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186. — Qwen2.5-Coder Technical Report
Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D Manning, and Chelsea Finn. 2023. Detectgpt: Zero-shot machine-generated text detection using probability curvature. In International conference on machine learning, pages 24950–24962. PMLR.
Daniil Orel, Dilshod Azizov, Indraneil Paul, Yuxia Wang, Iryna Gurevych, and Preslav Nakov. 2026. SemEval-2026 task 13: Detecting machine-generated code with multiple programming languages, generators, and application scenarios. In Proceedings of the 20th International Workshop on Semantic Evaluatio…

Updated 31 October 2026

This page in Markdown → https://md.research.elamur.ai/free/en/papers/teja-2026-codedet-nits-semeval-2026-task/