license: mit
📜 ProofWalaDataset
The ProofWalaDataset is a multilingual dataset of formal theorem proving traces collected from multiple interactive theorem prover (ITP) ecosystems. It provides a structured view of proof steps, goals, hypotheses, and theorem names from diverse mathematical and program verification libraries.
This dataset is intended for researchers and practitioners working on:
- Automated theorem proving
- Formal code generation
- Machine learning for logic
- Proof step prediction
- Multi-language transfer in formal systems
📂 Dataset Structure
The dataset is organized into the following ITP families:
lean/coq/GeoCoq/math-comp/multilingual/(cross-formal-language hybrid)
Each family includes standard splits:train/, test/, and eval/, each containing multiple JSON files.
Each JSON file contains a top-level key: "training_data" with a list of proof records.
🔍 Each record contains
| Field | Description |
|---|---|
proof_id |
Unique identifier for the proof trace |
goal_description |
Optional natural language description of the proof |
start_goals |
List of starting goals (each with goal and hypotheses) |
end_goals |
Final goals after applying proof steps |
proof_steps |
List of applied proof tactics (inv, rewrite, etc.) |
simplified_goals |
Simplified representations of goals (if any) |
all_useful_defns_theorems |
Set of useful definitions or theorems (static analysis) |
addition_state_info |
Optional additional metadata about the proof context |
file_path |
Source file where the proof appears |
project_id |
The ITP project or repository path (e.g., CompCert) |
theorem_name |
Name of the theorem being proved |
For convenience, structured fields such as start_goals[*].goal, start_goals[*].hypotheses, end_goals[*].goal, and end_goals[*].hypotheses are exposed directly through the Croissant metadata.
🧠 Use Cases
- Pretraining and finetuning LLMs for formal verification
- Evaluating proof search strategies
- Building cross-language proof translators
- Fine-grained proof tactic prediction
📊 Format
- Data format: JSON
- Schema described via Croissant metadata (
croissant.json) - Fully validated using mlcroissant