TeleAI at SemEval-2026 Task 13: Data-Centric Full-Parameter Fine-Tuning with Multi-Level Ensembling for Generated Code Detection

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Software Development & Engineering · Depth: Expert, quick

Summary

TeleAI's system achieved top rankings in SemEval-2026 Task 13 for multi-lingual and distribution-shift code generation detection, securing 1st place in Subtasks A and B, and 2nd place in Subtask C. Their framework integrates data-centric analysis, full-parameter model adaptation, and multi-level ensemble learning. The team first analyzed label and length distributions, applying repeated oversampling to address class imbalance. They then optimized prompts in a data-driven manner for improved inference stability. Based on the Qwen3-30B-A3B-Instruct model, TeleAI conducted full-parameter fine-tuning with diverse training configurations. Multiple checkpoints were integrated using soft voting, hard voting, logits-based voting, and LightGBM stacking, demonstrating substantial improvements over zero-shot baselines and consistent gains from ensemble strategies for robust detection.

Key takeaway

For Machine Learning Engineers developing robust code generation detection systems, integrating data-centric strategies with advanced ensembling is crucial. You should address class imbalance through oversampling and optimize prompts for inference stability. Fine-tuning powerful models like Qwen3-30B-A3B-Instruct and employing multi-level ensemble techniques, including soft voting and LightGBM stacking, will significantly enhance detection accuracy and resilience against distribution shifts. This approach can improve your system's performance in challenging multi-lingual environments.

Key insights

Combining data-centric analysis, full-parameter fine-tuning, and multi-level ensembling significantly improves code generation detection.

Principles

• Address class imbalance with repeated oversampling.
• Optimize prompts data-drivenly for stability.
• Ensemble diverse model checkpoints for robustness.

Method

The method involves analyzing data distributions, oversampling for imbalance, data-driven prompt optimization, full-parameter fine-tuning of Qwen3-30B-A3B-Instruct, and multi-level ensembling via various voting and stacking methods.

In practice

• Apply repeated oversampling for imbalanced datasets.
• Use LightGBM for ensemble stacking.
• Fine-tune Qwen3-30B-A3B-Instruct for code tasks.

Topics

• Code Generation Detection
• Full-Parameter Fine-Tuning
• Multi-Level Ensembling
• Data-Centric AI
• Qwen3-30B-A3B-Instruct
• SemEval-2026

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Code Is More Than Text: Uncertainty Estimation for Code Generation — Code generation uncertainty estimation requires specialized methods accounting for code's unique properties, outperforming natural language-derived baselines.
• Build an LLM from Scratch 6: Finetuning for Classification — Fine-tuning pre-trained LLMs for specific classification tasks is efficient and effective, even with smaller models.
• An Empirical Study of Multi-Generation Sampling for Jailbreak Detection in Large Language Models — Multi-generation sampling significantly improves jailbreak detection in LLMs compared to single-output evaluation.
• Exploring LLM-based Verilog Code Generation with Data-Efficient Fine-Tuning and Testbench Automation — Multi-agent LLMs can automate testbench generation, creating high-quality data for efficient Verilog code generation.
• Towards Functional Correctness of Large Code Models with Selective Generation — Code hallucination in LLMs can be controlled by a selective generator using fuzzing-generated unit tests and "α-code entailment".
• LinguIUTics at PsyDefDetect: Iterative Imbalance-Aware Fine-tuning of Qwen3-8B for Psychological Defense Mechanism Classification — Iterative imbalance-aware fine-tuning of LLMs, combined with augmentation and post-processing, significantly improves rare-class NLP performance.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.