MathlibGraph / README.md
Xinze-Li-Moqian's picture
Update dataset card with hold-out experiment results
8c70646 verified
metadata
license: apache-2.0
task_categories:
  - graph-ml
language:
  - en
tags:
  - mathematics
  - lean4
  - mathlib
  - dependency-graph
  - formal-verification
  - network-analysis
size_categories:
  - 100K<n<1M

MathlibGraph: The Multinetwork of Mathlib

Dependency graph of Mathlib (commit 534cf0b, 2 Feb 2026), the largest formal mathematics library for Lean 4 (v4.28.0-rc1).

Three dependency layers (declarations, modules, namespaces), each with nodes, edges, and precomputed network metrics.

Quick Stats

Declarations Modules Namespaces (k=2)
Nodes 308,129 7,564 10,097
Edges 8,436,366 20,881 332,081 (weighted)
DAG depth 83 154 7 (after SCC condensation)
Louvain modularity 0.478 0.610 0.270
Synthesized edges 74.2%
In cycles 5,732 nodes 0 (DAG) 6,055 nodes (38 SCCs)

Files

Raw Data (v1)

File Rows Description
mathlib_edges.csv 8,436,366 Declaration dependency edges
mathlib_nodes.csv 317,655 Declarations before deduplication
nodes.csv / edges.csv 633K / 10.9M Full environment (Lean + Std + Mathlib)
mechanisms.ndjson Lean language mechanism extractions
tactic_usage.ndjson Per-declaration tactic usage profiles

Three-Layer Graphs (v2)

Each layer has nodes.csv, edges.csv, and metrics.csv:

Folder Nodes Edges Metrics Description
v2/declaration/ 308,129 (use root mathlib_edges.csv) 308,129 x 11 Theorems, definitions, and other named constants
v2/module/ 7,564 20,881 7,564 x 10 Source files linked by import statements
v2/namespace/ 10,097 332,081 10,097 x 11 Depth-2 dotted-name prefixes, weighted edges

v2/summary.json contains headline statistics for all three levels.

Quick Start

from datasets import load_dataset

# Declaration level
decl = load_dataset("MathNetwork/MathlibGraph",
                    data_files="v2/declaration/metrics.csv", split="train").to_pandas()
edges = load_dataset("MathNetwork/MathlibGraph",
                     data_files="mathlib_edges.csv", split="train").to_pandas()

# Module level
mod_nodes = load_dataset("MathNetwork/MathlibGraph",
                         data_files="v2/module/nodes.csv", split="train").to_pandas()
mod_edges = load_dataset("MathNetwork/MathlibGraph",
                         data_files="v2/module/edges.csv", split="train").to_pandas()

# Namespace level
ns_nodes = load_dataset("MathNetwork/MathlibGraph",
                        data_files="v2/namespace/nodes.csv", split="train").to_pandas()
ns_edges = load_dataset("MathNetwork/MathlibGraph",
                        data_files="v2/namespace/edges.csv", split="train").to_pandas()

Schema

v2/declaration/nodes.csv

Deduplicated from mathlib_nodes.csv (317,655 to 308,129 rows; 9,526 @[to_additive] mirrors removed).

Column Type Description
name str Fully qualified declaration name
kind str theorem, definition, abbrev, inductive, constructor, opaque, axiom, or quotient
module str Parent namespace (null for 30,944 Lean core declarations)

v2/declaration/metrics.csv

All columns from nodes.csv plus:

Column Type Description
namespace_depth2 str First 2 dot-separated components (e.g., Mathlib.Algebra)
namespace_depth3 str First 3 dot-separated components
in_degree int Declarations that depend on this one
out_degree int Declarations this one depends on
pagerank float PageRank (alpha=0.85)
betweenness float Betweenness centrality (k=500, seed=42)
community_id int Louvain community
dag_layer int Topological depth; -1 for 5,732 nodes in cycles
file_module str Source file module (from Lean environment, 278K coverage)
is_tactic_proof bool Whether the proof uses tactics (78,315 declarations)
tactic_count int Number of tactics used in the proof
top_tactic str Most frequently used tactic in this proof
is_instance bool Whether this is a typeclass instance (26,415 declarations)
instance_class str Which typeclass this instance implements
is_coercion bool Whether this is a coercion (241 declarations)
to_additive_pair str Name of the corresponding additive/multiplicative variant
def_height float Definitional height in the kernel (42,935 declarations)

v2/module/nodes.csv

Column Type Description
module str Dotted module name (e.g., Mathlib.Algebra.Group.Defs)
decl_count int Declarations defined in this source file

v2/module/edges.csv

Column Type Description
source str Importing module
target str Imported module
is_exported bool Whether this is a public import (20,699 of 20,881)

v2/module/metrics.csv

Column Type Description
module str Module name
decl_count int Declarations in this module (full decl_module mapping, 499K coverage)
in_degree int Modules that import this one
out_degree int Modules this one imports
pagerank float PageRank on module import graph
betweenness float Betweenness centrality (exact)
dag_layer int Topological depth in module DAG
community_id int Louvain community
cohesion float Fraction of declaration edges staying within module
import_utilization_median float Median fraction of imported declarations used

v2/namespace/nodes.csv

Column Type Description
namespace str Depth-2 namespace (e.g., Mathlib.Algebra)
decl_count int Declarations in this namespace
in_cycle bool True if this namespace is in a strongly connected component
scc_id int SCC identifier (-1 if not in a cycle)

v2/namespace/edges.csv

Column Type Description
source str Source namespace
target str Target namespace
weight int Number of declaration-level edges between these namespaces

v2/namespace/metrics.csv

Column Type Description
namespace str Depth-2 namespace
decl_count int Declarations in this namespace
in_degree int Unweighted in-degree
out_degree int Unweighted out-degree
edge_weight_sum int Total declaration edges involving this namespace
pagerank float Weighted PageRank (alpha=0.85)
betweenness float Weighted betweenness (k=300, seed=42)
community_id int Louvain community (weighted undirected)
cross_ns_ratio float Fraction of edges crossing namespace boundaries
in_cycle bool In a strongly connected component (6,055 of 10,097)
scc_id int SCC identifier (-1 if acyclic)

Methodology

  • Extraction: lean4export (nodes) + lean-training-data (edges) + importGraph (module graph) + jixia (metadata)
  • Deduplication: by name, 317,655 to 308,129 rows (9,526 @[to_additive] mirrors)
  • Self-loops: 4,755 constructor self-references filtered
  • PageRank: alpha=0.85, max_iter=100, tol=1e-6
  • Betweenness: declaration k=500, namespace k=300, module exact; seed=42
  • Communities: Louvain, resolution=1.0, random_state=42, undirected projection
  • DAG layers: Kahn's algorithm; cycle nodes get layer=-1; namespace graph condensed via SCC
  • Namespace cycles: 6,055 of 10,097 namespaces in 38 SCCs (largest: 5,899 nodes)

Hold-Out Experiments

We validate the premise retrieval results with two levels of hold-out experiments to assess information leakage. In the original experiment, all network features (degree, PageRank, betweenness, community, DAG layer) are precomputed on the full graph. Hold-out experiments recompute features on a reduced graph to test whether full-graph computation inflates AUC.

Edge-level hold-out

  • Randomly remove 20% of edges, recompute all features on remaining 80% graph
  • Tests whether full-graph feature computation inflates AUC

Declaration-level hold-out

  • Split theorems 80/20, remove ALL edges (incoming and outgoing) of test theorems from training graph
  • Test theorems have zero degree in training graph
  • Simulates the real scenario: predicting premises for a brand new theorem

Results (Split 1, seed=42)

Method AUC (original) AUC (edge hold-out) AUC (decl hold-out)
Random 0.499 0.497 0.500
Same module 0.563 0.562 0.563
Same namespace 0.590 0.590 0.592
Same community 0.768 0.755 0.500
Network features 0.991 0.988 0.978
All features 0.994 0.992 0.984

The community feature drops to random (0.500) in declaration-level hold-out because test declarations become isolated nodes with unique community IDs. Despite this, the combined network feature model retains AUC=0.978, confirming that degree, PageRank, and DAG position carry genuine predictive signal independent of information leakage.

Full 5-split results with mean and std are in experiments/holdout_decl_level.json (updated incrementally).

Experiment Files

File Description
experiments/premise_retrieval_results.json Original experiment: 6 methods, 4 metrics, 95% CI
experiments/premise_retrieval_hard_negatives.json Hard negatives (same-community) variant
experiments/holdout_edge_level.json Edge-level hold-out (1 split)
experiments/holdout_decl_level.json Declaration-level hold-out (5 splits, incremental)

License

Apache 2.0