Guardado en:
Detalles Bibliográficos
Autores principales: Amballa, Avinash, Saluru, Durga Sandeep, Akkinapalli, Gayathri, Sureddy, Abhishek, Sureddy, Akshay Kumar
Formato: Preprint
Publicado: 2024
Materias:
Acceso en línea:https://arxiv.org/abs/2412.00074
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912138648354816
author Amballa, Avinash
Saluru, Durga Sandeep
Akkinapalli, Gayathri
Sureddy, Abhishek
Sureddy, Akshay Kumar
author_facet Amballa, Avinash
Saluru, Durga Sandeep
Akkinapalli, Gayathri
Sureddy, Abhishek
Sureddy, Akshay Kumar
contents Large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning and text generation. However, these models can inadvertently generate unsafe or biased responses when prompted with problematic inputs, raising significant ethical and practical concerns for real-world deployment. This research addresses the critical challenge of developing language models that generate both helpful and harmless content, navigating the delicate balance between model performance and safety. We demonstrate that incorporating safety-related instructions during the instruction-tuning of pre-trained models significantly reduces toxic responses to unsafe prompts without compromising performance on helpfulness datasets. We found Direct Preference Optimization (DPO) to be particularly effective, outperforming both SIT and RAFT by leveraging both chosen and rejected responses for learning. Our approach increased safe responses from 40$\%$ to over 90$\%$ across various harmfulness benchmarks. In addition, we discuss a rigorous evaluation framework encompassing specialized metrics and diverse datasets for safety and helpfulness tasks ensuring a comprehensive assessment of the model's capabilities.
format Preprint
id arxiv_https___arxiv_org_abs_2412_00074
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle Safe to Serve: Aligning Instruction-Tuned Models for Safety and Helpfulness
Amballa, Avinash
Saluru, Durga Sandeep
Akkinapalli, Gayathri
Sureddy, Abhishek
Sureddy, Akshay Kumar
Computation and Language
Large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning and text generation. However, these models can inadvertently generate unsafe or biased responses when prompted with problematic inputs, raising significant ethical and practical concerns for real-world deployment. This research addresses the critical challenge of developing language models that generate both helpful and harmless content, navigating the delicate balance between model performance and safety. We demonstrate that incorporating safety-related instructions during the instruction-tuning of pre-trained models significantly reduces toxic responses to unsafe prompts without compromising performance on helpfulness datasets. We found Direct Preference Optimization (DPO) to be particularly effective, outperforming both SIT and RAFT by leveraging both chosen and rejected responses for learning. Our approach increased safe responses from 40$\%$ to over 90$\%$ across various harmfulness benchmarks. In addition, we discuss a rigorous evaluation framework encompassing specialized metrics and diverse datasets for safety and helpfulness tasks ensuring a comprehensive assessment of the model's capabilities.
title Safe to Serve: Aligning Instruction-Tuned Models for Safety and Helpfulness
topic Computation and Language
url https://arxiv.org/abs/2412.00074