Update README.md

README.md

@@ -19,15 +19,24 @@ This repository hosts a single PyTorch Geometric dataset file used for the TAG-D
 
 ## Contents
 
-- `data.pt`  
-  A serialized object with the following keys:
-  - `data` (PyG `Data` object with collated node features, edges and labels)
-  - `slices` (index information for reconstructing individual graphs)
-  - `train_idx` (tensor of graph indices for training)
-  - `val_idx` (tensor of graph indices for validation)
-  - `test_idx` (tensor of graph indices for testing)
-
-## Basic usage
+- **`data.pt`**  
+  A preprocessed graph dataset containing:
+  - `data` — collated PyG `Data` object storing all node features, edges, and labels  
+  - `slices` — indexing information for reconstructing individual graphs  
+  - `train_idx`, `val_idx`, `test_idx` — fixed graph-level splits
+
+## Dataset Description
+
+- Contains the **first 10,000 proofs** from the ~45,000-theorem Metamath database  
+- Each example is a **directed acyclic graph** (DAG)  
+- **Node features:** 768-dimensional **CodeBERT embeddings** of Metamath statements  
+- **Labels:** the **theorem required to justify each node**  
+- **Conclusion masking:** the conclusion node’s embedding is zeroed out so the model must predict the final logical step directly from the graph structure and the other nodes  
+- **Rare labels** (<=5 occurrences) are collapsed into a single UNK class  
+- **All theorem statements** (not just proofs) are included in training, since the model must treat theorems themselves as prior knowledge  
+- Under the Metamath proof language, any theorem used to justify a step always has an index <= the theorem being proved. So a later theorem never appears in an earlier proof
+
+## Basic Usage
 
 ```python
 import torch