A Low-Cost Human-in-the-Loop Investigation of Toxicity on GitHub at Scale

2026-07-28

AI Summary

This work addresses the challenges posed by the implicit and context-dependent nature of toxic discourse on GitHub, which hinders large-scale annotation and limits the generalizability of existing research. To overcome this, the authors propose a low-cost human-in-the-loop annotation framework that leverages a local small language model to generate toxicity predictions along with interpretable event-category scores. A lightweight random forest verifier then identifies high-risk samples for human review. This approach substantially reduces annotation costs while outperforming confidence-based and multi-model baselines in both efficiency and accuracy. Applying the framework, the study successfully annotated 124,000 GitHub issue and pull request discussions, revising conclusions from prior small-scale studies and systematically uncovering the prevalence, characteristics, and dynamic evolution of toxic behavior in open-source communities.

Abstract

Toxic interactions in open source discussions can alienate contributors and threaten project sustainability, yet prior empirical studies of GitHub toxicity have been limited in scale, raising questions about their generalizability. Scaling up is difficult because toxicity on GitHub is often implicit and context-dependent, making both fully manual annotation and LLM-based labeling unreliable. We present a human-in-the-loop (HITL) annotation methodology that makes large-scale, domain-calibrated toxicity labeling practical. A single call to a small, local LLM produces both a toxicity prediction and a set of interpretable event category scores. A lightweight Random Forest validator then uses those scores to flag the small subset of conversations most likely to be mislabeled, directing human review only where it is needed. The validator outperforms confidence-based and multi-LLM baselines while adding low annotation cost. We apply this pipeline to over 124,000 GitHub issue and pull request conversations. Using the resulting dataset, we evaluate key findings from prior small-scale research, confirming some and qualifying others, and present new insights into the prevalence, characteristics, and dynamics of toxicity across diverse open source projects.

Problem

• toxicity
• GitHub
• human-in-the-loop
• annotation
• open source

Innovation

• human-in-the-loop
• toxicity detection
• lightweight Random Forest validator
• interpretable event categories
• large-scale annotation