DebateLLM — Multi-Label Fallacy Detection

Fine-tuned DeBERTa-base model for multi-label logical fallacy detection in debate transcripts.

Model Details

• Base model: microsoft/deberta-base
• Task: Multi-label text classification
• Fallacy classes: Ad Hominem, Straw Man, Weak Rebuttal, Unsupported Claim, Hasty Generalization, Circular Reasoning, Appeal to Emotion, False Dilemma
• Training data: 9,400 debate-style examples
• Validation F1-macro: 1.0
• Hard OOD test: 13/15

Purpose

Built as the LocalLLMClient fallback for the Explainable Debate Intelligence System (EDIS) — a GSoC 2026 proposal for DebateAI under AOSSIE.