DeEscalWild: A Real-World Benchmark for Automated De-Escalation Training with SLMs

2026-03-20

AI Summary

This study addresses the limitations of existing de-escalation training—namely, its lack of scalability and realism—and the practical challenges of deploying large language models on edge devices, compounded by the scarcity of domain-specific data for smaller models. To bridge this gap, the authors introduce DeEscalWild, the first high-quality dialogue dataset derived from real-world police-civilian interaction videos. Through human-in-the-loop filtering, 1,500 high-fidelity scenarios were curated from an initial pool of 5,000 samples. A lightweight model (Qwen2.5-3B-Instruct) was then fine-tuned on this dataset. Experimental results demonstrate that this domain-specialized small model significantly outperforms baseline approaches across automatic metrics (ROUGE-L, BLEU-4, METEOR, BERTScore), perceived realism, and human evaluations, even surpassing Gemini 2.5 Flash. These findings validate the feasibility of delivering effective, real-time de-escalation training with minimal computational overhead.

Abstract

Effective de-escalation is critical for law enforcement safety and community trust, yet traditional training methods lack scalability and realism. While Large Language Models (LLMs) enable dynamic, open-ended simulations, their substantial computational footprint renders them impractical for deployment on the lightweight, portable hardware required for immersive field training. Small Language Models (SLMs) offer a viable real-time alternative but suffer from a critical scarcity of high-quality, domain-specific training data. To bridge this gap, we present DeEscalWild, a novel benchmark dataset curated from a multi-stage pipeline of in-the-wild police-civilian interactions extracted from publicly available video repositories. Starting with 5,000 raw inputs, we employed a rigorous hybrid filtering process combining human-in-the-loop verification with LLM-as-a-Judge evaluation to distill 1,500 high-fidelity scenarios. The resulting corpus comprises 285,887 dialogue turns, totaling approximately 4.7 million tokens. Extensive experiments demonstrate that SLMs fine-tuned on this data significantly outperform their base counterparts across ROUGE-L, BLEU-4, METEOR, BERTScore, Realism Score, and human evaluation metrics. Notably, our fine-tuned Qwen 2.5 (3B-Instruct) surpasses the general-purpose Gemini 2.5 Flash model when evaluated under equivalent conditions, demonstrating that domain-optimized SLMs can achieve superior performance with a fraction of the computational cost. This work establishes the foundational infrastructure for accessible, low-latency, and privacy-preserving officer training systems at the edge. We publicly release our code(https://github.com/Hasebul/DeEscalWild-Benchmark-Framework) and dataset(https://doi.org/10.7910/DVN/CWMCZI).

Problem

• de-escalation training
• Small Language Models
• real-world benchmark
• domain-specific data
• law enforcement

Innovation

• Small Language Models
• De-escalation Training
• Real-World Benchmark
• Edge Deployment
• Domain-Specific Fine-tuning