Jack11211/REVERSE

REVERSE数据集是一个用于强化证据验证和搜索的代理图像地理定位数据集，支持训练和评估视觉语言模型在图像地理定位任务上。它采用多轮工具增强推理方法，包括缩放、图像搜索和文本搜索工具。数据集包含约10,000张训练图像（808MB），并分为监督微调（SFT）和强化学习（RL）两个主要部分：SFT部分包括冷启动训练和验证样本，RL部分包括简单课程和完整课程训练样本，以及验证数据。此外，还提供离线图像搜索和文本搜索缓存。数据集结构清晰，图像路径在parquet文件中为相对路径，使用时需设置DATA_ROOT环境变量指向本地下载路径。

The REVERSE dataset is designed for reinforcing evidence verification and search for agentic image geolocation. It supports training and evaluating vision-language models on image geolocation using multi-turn tool-augmented reasoning, such as zoom, image search, and text search. The dataset includes approximately 10,000 training images (808MB) and is structured into supervised fine-tuning (SFT) and reinforcement learning (RL) components: SFT includes coldstart training and validation samples, while RL includes easy curriculum and full curriculum training samples, along with validation data. Additionally, it provides offline image search and text search caches. The dataset structure is organized with relative image paths in parquet files, requiring the DATA_ROOT environment variable to be set to the local download path for usage.