zs0506/SVE-Math-DeepSeek-7B

SVE-Math-DeepSeek+-7B

🤖 SVE-Math-DeepSeek+-7B is a fine-tuned Multi-modal Large Language Model (MLLM) built upon SVE-Math-DeepSeek-7B, further enhanced using GeoPeP, a perception-oriented dataset of 200K high-quality geometry image-text pairs.

This model is released as part of our project:
📘 "Open Eyes, Then Reason: Fine-grained Visual Mathematical Understanding in MLLMs (SVE-Math)"
🔗 Paper & Code: github.com/AI4Math-ShanZhang/SVE-Math

---

🔍 Key Features

• 💡 Designed to improve visual perception in mathematical diagrams.
• 📊 Fine-tuned on high-quality perception-oriented dataset (100K diagram-caption + 100K conversation).
• 🧠 GeoPeP explicitly structs diagrams into shapes, attributes, locations and relationships.
• ⚙️ Systematic investigation explores how visual perception impacts mathematical reasoning in MLLMs.

---

💬 Inference

You can refer to the official inference code and setup from our GitHub repo:
👉 https://github.com/AI4Math-ShanZhang/SVE-Math