Human-Like-DPO-Dataset

Summary

A DPO training dataset of 10,884 samples across 256 topics, each containing a conversational question, a human-like response (natural and engaging), and a formal response (structured and professional). It was created as part of research on improving conversational fluency in LLMs, aiming to make AI interactions feel more conversational and emotionally intelligent without sacrificing accuracy. The associated paper was accepted to the AAAI-26 Workshop on Personalization in the Era of Large Foundation Models (PerFM).

Keywords

dpo human-like conversational-ai preference-optimization emotional-intelligence llm-fine-tuning aaai-26

Topics

NLP / Conversational AI

Research notes

• Paper: arxiv 2501.05032 'Enhancing Human-Like Responses in Large Language Models'. License is Llama 3 license. Dataset includes multiple file parts (part1, part2) with casual conversations and general instructions.