REVERSE Dataset

Dataset for REVERSE (Reinforcing Evidence Verification and Search for Agentic Image Geolocation).

This dataset supports training and evaluating vision-language models on image geolocation using multi-turn tool-augmented reasoning (zoom, image search, text search).

Structure

REVERSE/
├── images/                          # Training images (808MB, ~10k images)
├── sft/
│   ├── train_sft_coldstart_v5.parquet   # SFT coldstart train (2,217 samples)
│   └── val_sft_coldstart_v5.parquet     # SFT coldstart val (116 samples)
├── rl/
│   ├── train_rl_v5_easy.parquet         # RL Phase 1 easy curriculum (4,080 samples)
│   ├── train_rl_v5_full.parquet         # RL Phase 2 full (9,855 samples)
│   └── val_rl.parquet                   # RL validation
└── rl_cache/
    ├── image_search_cache.parquet       # Offline image search cache
    └── text_search_cache.parquet        # Offline text search cache

Usage

from datasets import load_dataset

# Load SFT data
sft = load_dataset("Jack11211/REVERSE", data_files="sft/train_sft_coldstart_v5.parquet")

# Load RL data
rl = load_dataset("Jack11211/REVERSE", data_files="rl/train_rl_v5_easy.parquet")

Image paths in the parquet files are relative to the dataset root. Set DATA_ROOT to the local download path when training:

export DATA_ROOT=/path/to/REVERSE

Code

Training and evaluation code: REVERSE (GitHub)