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

DeEscalWild is a large-scale, real-world benchmark dataset designed to facilitate automated de-escalation and crisis intervention training using Small Language Models (SLMs). Extracted from publicly available \"in-the-wild\" social media video repositories (including YouTube, TikTok, and Facebook), the dataset captures the unstructured, volatile nature of real-world police-civilian interactions, encompassing speech disfluencies, emotional outbursts, and chaotic multi-party environments. Unlike chitchat or standard task-oriented dialogue systems, DeEscalWild focuses on reasoning-centric de-escalation. It is designed to train and evaluate models on their ability to infer latent mental states, anticipate escalation triggers, and deploy communicative actions that actively lower tension.