Human-Like-Qwen2.5-7B-Instruct Overview

This model is a specialized fine-tuned version of the Qwen2.5-7B-Instruct base model, developed by HumanLLMs. Its primary objective is to produce more human-like and conversational responses, enhancing natural language understanding and emotional intelligence in interactions. The development process, detailed in the research paper "Enhancing Human-Like Responses in Large Language Models" (accepted to AAAI-26 PerFM Workshop), involved fine-tuning with both Low-Rank Adaptation (LoRA) and Direct Preference Optimization (DPO).

Key Training Details

• Base Model: Qwen2.5-7B-Instruct
• Training Methods: LoRA and DPO
• Dataset: A synthetic dataset comprising approximately 11,000 samples across 256 diverse topics, generated using LLaMA 3 models. This dataset includes both human-like and formal responses.
• Hardware: Trained on 2x NVIDIA A100 (80 GB) GPUs for about 2 hours and 15 minutes.

Performance Insights

While the fine-tuning prioritizes human-like responses, benchmark results show some trade-offs compared to the base Qwen-2.5-7B-Instruct model. For instance, it shows a slight decrease in overall average performance but demonstrates improvements in specific areas like GPQA (+1.01) and MMLU-PRO (+1.24), indicating enhanced reasoning and knowledge recall in certain contexts. The model is part of a series of "Human-Like" models, including variants based on Llama-3-8B and Mistral-Nemo-Instruct.

Ideal Use Cases

This model is particularly well-suited for applications where the naturalness and conversational quality of AI responses are paramount. Consider using it for:

• Chatbots and virtual assistants requiring highly engaging and empathetic interactions.
• Creative writing or dialogue generation where human-like conversational flow is crucial.
• Interactive storytelling or role-playing scenarios that benefit from nuanced emotional intelligence.