Published June 3, 2026 | Version Version 1
Preprint Open

TAME: A 20,000-Generation Controlled Evaluation of Toxicity Amplification and Intersectional Bias in Locally Deployed Large Language Models

Authors/Creators

  • 1. Independent Researcher

Description

TAME-Research

TAME-Research is a reproducible evaluation pipeline for measuring toxicity amplification in local large language models. The repository preserves the prompt sets, raw generations, scored outputs, analysis tables, and combined statistical summaries used for the TAME study.

Paper Title

TAME: Toxicity Amplification Measurement and Evaluation for Local LLM Auditing

Project Scope

  • Total analyzed generations: 20,000
  • Models: 5 local Ollama models
  • Series 1 models: mistral, llama3, phi3:mini-4k
  • Series 2 models: qwen3:8b, gemma3:4b
  • Experimental design: neutral, single-identity, and intersectional prompt variants scored with Detoxify and analyzed with paired and aggregate comparison metrics

Files

TAME_Research_Paper_Author_Aayush_Sevda.pdf

Files (1.8 MB)

Name Size
md5:5633580b9a51b3265bf94671685b9777
760.3 kB Download
md5:e2f58a7fb678c64f0d777248f4201dcd
994.9 kB Preview Download

Additional details

Dates

Submitted
2026-06-03
Initial public release of the TAME benchmark study and accompanying repository.

Software

Repository URL
https://github.com/AayushSevda/TAME-Research
Programming language
Python , YAML , CSV
Development Status
Active

References