Information
# Asking like Socrates: RS-EoT-4K Dataset
[** Project Website**](https://geox-lab.github.io/Asking_like_Socrates/) | [** GitHub Repository**](https://github.com/GeoX-Lab/Asking_like_Socrates) | [** Paper (ArXiv)**](https://arxiv.org/abs/2511.22396) | [** Model (RS-EoT-7B)**](https://huggingface.co/ShaoRun/RS-EoT-7B)
## Introduction
**RS-EoT-4K** is a multimodal instruction-tuning dataset specifically designed to instill **Evidence-of-Thought (EoT)** reasoning capabilities into Vision-Language Models (VLMs) for Remote Sensing (RS) tasks.
This dataset was introduced in the paper **"Asking like Socrates: Socrates helps VLMs understand remote sensing images"**. To address the widespread "Glance Effect" in RS tasks—where models narrate reasoning without genuine visual inspection—we propose **SocraticAgent**, a self-play multi-agent system. SocraticAgent synthesizes reasoning traces via alternating cycles of linguistic deduction (Reasoner) and visual inspection (Perceiver), creating the RS-EoT-4K dataset.
This dataset is used for the **SFT Cold-Start** stage of the **RS-EoT-7B** model.
## Dataset Statistics
The RS-EoT-4K dataset comprises approximately **4.3K** high-quality reasoning samples covering multiple modalities (RGB, Infrared, SAR). The data is synthesized from the following sources:
| Data Source | Modality | Count |
| :--- | :--- | :--- |
| **FIT-RS** | RGB | ~1.9K |
| **VRSBench** | RGB | ~1.1K |
| **EarthVQA** | RGB | ~0.6K |
| **RSVQA** | RGB | ~0.3K |
| **Drone Vehicle** | RGB & Infrared | ~0.2K |
| **SARLang-1M** | SAR | ~0.2K |
| **Total** | **RGB/INF/SAR** | **~4.3K** |
## Methodology: SocraticAgent
The dataset was generated using **SocraticAgent**, a multi-agent framework inspired by the Socratic Method:
1. **Reasoner (Text-only):** Poses perceptual questions to gather evidence based on the task query.
2. **Perceiver (Visual-expert):** Provides accurate visual descriptions in response to the Reasoner's questions.
3. **Verifier:** Validates the final answer against ground truth to ensure data quality.
## Usage
You can load this dataset directly using the Hugging Face \`datasets\` library.
\`\`\`python
import datasets
import random
# Load the dataset from Hugging Face
data = datasets.load_dataset("ShaoRun/RS-EoT-4K")
# Print dataset structure
print(data)
# Print a random sample
print(random.choice(data['train']))
\`\`\`\`
## Related Models
The model trained using this dataset (and further optimized via RL) allows for iterative evidence-seeking reasoning:
* **RS-EoT-7B**: [https://huggingface.co/ShaoRun/RS-EoT-7B](https://huggingface.co/ShaoRun/RS-EoT-7B)
## ️ Citation
If you find this dataset or our work useful in your research, please consider citing:
\`\`\`bibtex
Coming soon
\`\`\`