TrustGeoGen: Formal-Verified Data Engine for Trustworthy Multi-modal Geometric Problem Solving

🚩 News

• [2025/8/29] GeoTrust-train and GeoTrust-test dataset have been released, link.
• [2025/4/21] Code and data will be open-sourced soon.
• [2025/4/21] Upload paper and init project.

Data Description

Field information of GeoTrust-train:

{
    "_id": "question ID",
    "image_path": "image path",
    "problem_FL": "problem in formal language",
    "solution_FL": "solutino in formal language",
    "problem_pretty": "prettied problem using rule-based method",
    "solution_pretty": "prettied solution using rule-based method",
    "problem_NL": "step-by-step translated problem using LLM",
    "solution_NL": "step-by-step translated solution using LLM",
    "connection_thinking": "connection thinking generated by LLM using solution_NL as a scaffold",
    "proof_length": "number of reasoning steps",
    "used_premises_nums": "number of used premises in solution",
    "all_premises_nums": "total number of premises"
}

Field information of GeoTrust-test

{
    "_id": "question ID",
    "image_path": "image path",
    "tier": "tier of the question",
    "problem": "problem in natural language",
    "answer": "final answer"
}