Scholarlink

<p>Recent breakthroughs in Natural Language Processing (NLP) have caused a surge in the application of Large Language Models (LLMs) for a variety of tasks, including content generation, translation, and decision support. However, these models tend to re-represent and magnify human biases present in their training data, leading to socially unfair predictions. We study bias mitigation through fine-tuning and prompt manipulation in three open-source models— LLaMA2-7B, Mistral-7B, and Dolly-7B. We develop six different prompting schemes and generate an inference dataset to measure how they mitigate bias. We also fine-tune every model to encourage neutral representations with respect to four social factors, including gender, race, occupation, and religion. Post-finetuning evaluations show that the neutrality and fairness of model outputs are greatly improved. The findings demonstrate the interplay of prompt variation and fine-tuning in mitigating societal bias in LLMs, guiding the building of more fair and ethical language technologies.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Fine-tuning and prompting: a strategy for mitigating societal biases in large language models

  • 期刊论文

  • 发表日期：

    2026年2月12日

查看全文
  • Pradeep Kamboj,
  • Shailender Kumar

摘要

Recent breakthroughs in Natural Language Processing (NLP) have caused a surge in the application of Large Language Models (LLMs) for a variety of tasks, including content generation, translation, and decision support. However, these models tend to re-represent and magnify human biases present in their training data, leading to socially unfair predictions. We study bias mitigation through fine-tuning and prompt manipulation in three open-source models— LLaMA2-7B, Mistral-7B, and Dolly-7B. We develop six different prompting schemes and generate an inference dataset to measure how they mitigate bias. We also fine-tune every model to encourage neutral representations with respect to four social factors, including gender, race, occupation, and religion. Post-finetuning evaluations show that the neutrality and fairness of model outputs are greatly improved. The findings demonstrate the interplay of prompt variation and fine-tuning in mitigating societal bias in LLMs, guiding the building of more fair and ethical language technologies.