AquaLabs/Spatial-DPO-Dataset

这是一个用于训练语言模型进行3D到语音转换的数据集，特别为EchoLLaMA项目设计。数据集包含2000个样本，每个样本由来自3D图像分析的提示和两种类型的响应组成，分别是DeepSeek-V3-0324模型生成的高质量响应和LLaMA-3.2-1B-Instruct模型生成的基线响应。这种结构使得可以直接优化语言模型以生成详细且与上下文相关的3D场景描述。

This dataset is designed for training language models for 3D-to-Speech conversion, specifically for the EchoLLaMA project. It contains 2,000 samples, each consisting of a prompt derived from 3D image analyses and two types of responses: high-quality responses generated by DeepSeek-V3-0324 and baseline responses generated by LLaMA-3.2-1B-Instruct. This structure allows for Direct Preference Optimization (DPO) to fine-tune language models for generating detailed and contextually accurate descriptions of 3D scenes.