ShriNep@EEUCA 2026: RAKSHAK – Multi-Task DeBERTa with Rationale Distillation and Jigsaw-Augmented Training for Toxic Intent Classification

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Natural Language Processing · Depth: Expert, medium

Summary

The ShriNep@EEUCA 2026 team introduced RAKSHAK, a multi-task DeBERTa-v3-base system designed for the GameTox Shared Task at ACL 2026. This system classifies World of Tanks chat utterances into six fine-grained toxic intent categories, addressing challenges like severe class imbalance and scarce data for categories such as Threats (60 samples) and Extremism (24 samples). RAKSHAK integrates rationale distillation from Qwen2.5-14B, Supervised Contrastive Loss, and rare-class binary heads. Its training data is augmented with 16,225 samples from the Jigsaw Toxic Comment dataset and 100 LLM-generated extremism samples. RAKSHAK achieved a Macro F1 of 0.5883, placing 7th among 35 teams. A secondary system, M1, fine-tuned DeBERTa-v3-base with Focal Loss, reaching 0.5252 Macro F1. Ablation showed cross-domain transfer added +2.6 F1 points, and RAKSHAK's multi-task architecture added +3.7 points.

Key takeaway

For NLP Engineers building toxic content classifiers, consider multi-task DeBERTa-v3-base architectures with rationale distillation. To address severe class imbalance and scarce data, augment your training sets. Use cross-domain transfer from larger datasets like Jigsaw. Also, generate synthetic samples using LLMs for rare categories. This approach, demonstrated by RAKSHAK's 0.5883 Macro F1, significantly improves fine-grained toxic intent classification. It especially benefits underrepresented labels like Threats and Extremism.

Key insights

Multi-task learning with rationale distillation and data augmentation significantly improves toxic intent classification, especially for rare classes.

Principles

• Cross-domain data transfer boosts performance.
• Multi-task architectures enhance classification.
• Rationale distillation improves model understanding.

Method

RAKSHAK employs a multi-task DeBERTa-v3-base with Qwen2.5-14B rationale distillation, Supervised Contrastive Loss, and rare-class binary heads. It augments data with Jigsaw transfer and LLM-generated samples.

In practice

• Augment rare classes with LLM-generated data.
• Use cross-domain datasets for transfer learning.
• Implement multi-task learning for complex labels.

Topics

• Toxic Intent Classification
• Multi-task Learning
• DeBERTa-v3-base
• Rationale Distillation
• Data Augmentation
• Jigsaw Dataset

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• A Multitask Transformer for Offensive Language Detection and Target Identification in HateBR — Multitask learning improves offensive language detection consistency but may reduce fine-grained target attribution.
• EGTR-Review: Efficient Evidence-Grounded Scientific Peer Review Generation via Multi-Agent Teacher Distillation — Multi-agent teacher distillation enables efficient, evidence-grounded, and traceable scientific peer review generation by training a lightweight student model.
• Distilling Examples into Task Instructions: Enhanced In-Context Learning for Real-World B2B Conversations — Distilling ICL examples into structured instructions significantly improves performance and efficiency for complex B2B conversation classification.
• 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.
• mdok-style at SemEval-2026 Task 10: Finetuning LLMs for Conspiracy Detection — Data augmentation and self-training can effectively finetune LLMs for specialized text classification tasks with limited data.
• Multi-Stage Training for Abusive Comment Detection in Indic Languages — Multi-stage training with language preprocessing and model ensembles improves abusive comment detection while minimizing false positives.

Open in AIssential →

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