Graphs#

Release:: 6.1.2
Date:: Oct 28, 2026

The graph catalog, organized by category: each category page lists its graphs with statistics and download links.

Format conversion#

Graphs can be converted between the formats below without building an in-memory NetworkX graph: the conversion runs over an edge stream — a lazy iterator of (u, label, v) tuples in the TXT FROM LABEL TO order — with one streaming reader and one streaming writer per format, composed by convert_graph. This keeps peak memory at O(1) for the common case (an MTX directory to TXT or .g text), where routing through a networkx.MultiDiGraph costs several gigabytes on the largest graphs (#138: ~4 GiB of RAM for ~100 MB of input).

Formats#

Format	On disk	Node ids	Direction
`mtx`	a directory with one Boolean MatrixMarket file per label (the File structure layout)	non-negative integers	both
`txt`	one edge per line, `FROM LABEL TO`, optional quoting	strings	both
`rdf`	Turtle (the encoding below)	strings	both
`g`	FastMatrixCFPQ `.g` text: `u v label` lines plus auto-generated reverse edges; an indexed family `X_0, X_1, ...` collapses to `X_i` with the index as a fourth column	non-negative integers	write-only
`graph`	not on disk — an in-memory `networkx.MultiDiGraph`	whatever the graph holds	read-only

Memory characteristics#

Every reader and writer streams: one edge is held in memory at a time, so mtx → txt / g / rdf runs in O(1) RAM (the #138 case).
Writing mtx is a single pass over the source with one temporary file per label (O(1) RAM, O(E) disk); it requires integer node ids (digit strings are accepted, anything else is an error).
An rdf source is materialized: rdflib has no streaming parser, so reading RDF costs O(E) RAM. It is the only unavoidable materialization.

RDF encoding#

The RDF writer emits valid RDF 1.1 Turtle, one triple per line:

<urn:flpq:node:0> <urn:flpq:label:load_5> <urn:flpq:node:1> .

Nodes are IRIs urn:flpq:node:<id>, predicates IRIs urn:flpq:label:<label>, with the variable components percent-encoded; the writer streams line by line without an in-memory RDF store. The reader accepts this encoding and, for backwards compatibility, the legacy form that older versions of graph_to_rdf emitted — blank-node endpoints with a Literal predicate, which is not valid RDF 1.1 (predicates must be IRIs) and only round-tripped inside rdflib.

RDF is a set of triples: parallel edges with identical (u, label, v) collapse on write — a format-inherent limitation the legacy form shared. The other formats preserve parallel edges.

The NetworkX boundary#

NetworkX stays where it earns its place: the random graph generators, the adjacency-based utilities (add_reverse_edges, filter_edges, …), query materialization, and the existing graph_from_* / graph_to_* API, which builds or consumes a networkx.MultiDiGraph. The conversion path above never materializes one.

CSV removal#

The former CSV format (one edge per line, from to label) is removed in 6.0.0: it was the same line-per-edge format as TXT with only the column order different — a footgun, not a feature — and no dataset archive uses it (archives are MTX-only). Use txt instead; graph_from_csv / graph_to_csv are gone, and the pandas dependency that only they used is dropped.

Contents#

The per-category tables include one count column per grammar language applicable to that category, each headed Reachable pairs (or Reachable pairs / <grammar> in categories with several languages, such as RDF). Each cell shows the number of vertex pairs reachable with respect to the respective language (i.e. the number of pairs (u, v) for which some string in the language labels a path from \(u\) to \(v\)). The count depends on the language, not on a particular grammar: several grammars may generate the same language and yield the same count. A cell reading “not available” means the value has not been computed yet; an empty cell means the grammar does not apply to that graph. Each table also has a Size (MB) column: the download size of the graph’s archive, in megabytes.

The canonical grammar behind each column is documented once on the respective category page, in its “Canonical grammars” section — a grammar is canonical for the whole category, so the per-graph pages do not repeat it.

Source area	Number of graphs
C alias analysis	20
RDF	20
Java points-to graphs	21
Field-Sensitive Alias	10
Context-Sensitive Data-Flow	10
Data Provenance	18
Name Resolution	4
Biological graphs from UniProt	10