Researchers have developed two systems, RAKSHAK and M1, for the GameTox Shared Task at ACL 2026, focusing on classifying toxic intent in World of Tanks chat. RAKSHAK, the primary system, utilizes a DeBERTa-v3-base model enhanced with rationale distillation from Qwen2.5-14B, Supervised Contrastive Loss, and dedicated heads for rare classes. It also incorporates cross-domain transfer from the Jigsaw Toxic Comment dataset and LLM-generated samples for extremism. RAKSHAK achieved a Macro F1 of 0.5883, ranking 7th out of 35 teams, while the secondary system M1 achieved 0.5252 Macro F1. AI
IMPACT This research demonstrates advanced techniques for handling imbalanced and scarce data in toxic intent classification, potentially improving moderation tools.
RANK_REASON The item is an academic paper detailing a system for a specific NLP task. [lever_c_demoted from research: ic=1 ai=1.0]
- ACL 2026
- An et al. Reply
- DeBERTa-v3-base
- GameTox Shared Task
- Jigsaw Toxic Comment dataset
- Qwen2.5:14b
- RAKSHAK
- ShriNep@EEUCA 2026
- World of Tanks
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
