license: apache-2.0
language:
- en
- zh
pipeline_tag: image-text-to-text
tags:
- remote-sensing
- geospatial-reasoning
- qwen2.5-vl
- sft
- chain-of-thought
- ms-swift
LandAI-Base: Activating Geospatial Chain-of-Thought Reasoning
[Paper (Under Review)]
📖 Introduction
LandAI-Base is the foundational Supervised Fine-Tuned (SFT) model of the LandAI family. Built upon the Qwen2.5-VL-7B-Instruct architecture, it is specifically designed to activate domain-specific logical reasoning in Earth Observation tasks.
Unlike general-purpose multimodal models, LandAI-Base has been fine-tuned on a composite corpus of approximately 334,000 reasoning chains, including the novel Geo-Base-Thinking-14K dataset. This process instills the model with the "epistemological authority" of geography experts, enabling it to decompose complex spatial problems before engaging in visual recognition.
LandAI-Base serves two primary purposes:
- A robust baseline for geospatial reasoning tasks (Q&A, analysis).
- The "Cold Start" initialization for the LandAI-L1 model (trained via GRPO-L1).
🚀 Key Features
- Domain-Specific Cognitive Activation: Fine-tuned to simulate the reasoning patterns of geography experts, moving from rote memorization to logical deduction.
- High-Quality Training Data: Trained on a curated mix of:
- Geo-Base-Thinking-14K: ~14.7k distillations from geography entrance exams and textbooks.
- General Reasoning Corpus: Subsets from OpenR1-Math, OpenThoughts, and Chinese-Data-R1 to enhance mathematical and scientific logic.
- Strong Zero-Shot Performance: Significantly outperforms the vanilla Qwen2.5-VL-7B on geographic benchmark exams.
- MS-Swift Compatibility: Fully compatible with the ms-swift training framework.
📊 Performance Benchmarks
LandAI-Base demonstrates a substantial leap in reasoning capabilities compared to its backbone model. In the GeoTest2025 benchmark (derived from restricted 2025 National Postgraduate Entrance Examination questions), it achieves near-commercial performance.
| Model | GeoTest2025 (Geography) | AIME 2024 | HumanEval | MMMU pro |
|---|---|---|---|---|
| LandAI-Base-7B (Ours) | 93.3% | 16.7% | 66.4% | 44.7% |
| Qwen2.5-VL-7B (Baseline) | 46.7% | 3.3% | 67.3% | 41.2% |
| GPT-4o | 92.1% | 9.3% | 90.2% | 51.9% |
| Gemini 2.5 Pro | 98.3% | 92.0% | - | 71.2% |
📂 Dataset Composition
The explicit reasoning capability of LandAI-Base stems from its training data distribution:
| Dataset Source | Samples | Purpose |
|---|---|---|
| Geo-Base-Thinking-14K | ~14.7k | Domain-specific geospatial logic & knowledge |
| OpenR1-Math | ~96k | Mathematical reasoning infrastructure |
| OpenThoughts | ~114k | General scientific literacy (Physics/Chem/Bio) |
| Chinese-Data-R1 | ~110k | Linguistic nuance and logic bridging |
🛠️ Quick Start
LandAI-Base follows the standard Qwen2.5-VL architecture. You can use it for geospatial Question Answering or as a base for further RL training.