zelph Binaries Dataset

This dataset offers pre-compiled binary files (.bin) intended for use with zelph, a semantic network system and reasoning engine. These binaries are derived from large knowledge bases like Wikidata, optimized for fast loading and efficient querying.

Dataset Description

zelph binaries enable users to work with semantic networks without the need to import raw dumps (e.g., JSON files), which can take hours. Instead, these .bin files load in minutes, though they require substantial RAM.

Which file do you want?

Select by the RAM column, not the file size. A .bin roughly triples when it is loaded: the graph is reconstructed as adjacency maps and name tables rather than mapped from the file.

File	Nodes	Download	RAM to load	Load time
wikidata-20260309-all.bin	983,424,620	82 GiB	223.7 GiB	23m 23s
wikidata-20260309-all-P11260.bin	983,435,690	82 GiB	221.7 GiB	—
wikidata-20260309-all-pruned-medium.bin	114,477,445	9.0 GiB	25.9 GiB	2m 08s
wikidata-20260309-all-pruned-small.bin	26,533,048	2.2 GiB	6.0 GiB	25s
wikidata-20260309-all-pruned-small-P279.bin	2,005,552	0.21 GiB	0.6 GiB	2.2s

• -small is the one to begin with: it loads in 25 seconds and leaves room to work on a 16 GiB machine. On 8 GiB it fits, but not with much to spare.
• -small-P279 is not a network to work in but solely the class hierarchy: every P279 statement from -small and nothing beyond. It delivers answers to class-hierarchy queries precisely as -small does, on any system, within two seconds.
• -medium is the same network with people still in it, and requires a 32 GiB machine.
• wikidata-20260309-all.bin holds every statement from the Wikidata JSON dump that links two entities – dates, quantities, identifiers such as DOIs, and qualifiers are left out – formatted so that it loads in minutes rather than hours. It needs a machine built for it.
• wikidata-20260309-all-P11260.bin contains the complete dump along with the list item qualifiers associated with the disjoint union of statements – the statement and qualifier layer utilized during disjointness analysis. All other files include only direct triples.

Historic 2017 dumps (wikidata-20171227.bin, 44.6 GiB RAM, and its pruned variant at 3.8 GiB) are also available.

What the pruned variants drop, and what they keep

Removed, in the order they went – least missed first: the encyclopedia’s own plumbing (categories, templates, list and disambiguation pages, one item per integer and per calendar day), astronomical catalogues, sequence databases and chemistry, individual museum objects, publications and their citation graph, everything filed under an administrative entity, everything carrying a country, and – in -small only – people.

Retained, intentionally: the class hierarchy. Eliminating instances of a class never eradicates the class itself, so both pruned versions respond to class-level queries identically to one another. The disjointness query outlined under Working on the Wikidata Class Hierarchy returns the same 81 topmost culprits on -medium and on -small; they differ only in which individual items they still include.

A file that appears by itself: .pidx

The very first time you pose a transitive query (P279+, or SPARQL with wdt:P279+), zelph generates a <file>.bin.pidx.322 adjacent to the network. It stores the persisted transitive closure of P279 (subclass of).

It is a machine-local cache, not data, and it is deliberately not part of this dataset: it stores raw pairs in host byte order, so it would be wrong on a machine of different endianness, and it is validated against the exact network it was built from. Yours is built automatically in about fifteen seconds, and deleting it costs time and never information.

Regarding sizes, the methodology behind pruning and the approach to load parts of a file without loading the entire file, refer to https://zelph.org/binaries.

How to Use

1. Download the desired .bin file from this dataset.
2. In zelph interactive mode, load it with:

   .load /path/to/your-file.bin
   

3. For Wikidata files, switch the display language so that Q/P ids resolve:

   .lang wikidata
   

4. Run queries, define rules, perform inferences or run complete scripts (refer to Querying in zelph and Rules and Inference for further information).

If a file is larger than your RAM, you do not have to give up on it: zelph can load parts of a .bin — selected chunks, or a graph without its name maps — which is described under Partial Loading.

The full dump and -small are additionally provided in a sharded format, within the directories wikidata-20260309-all/ and wikidata-20260309-all-pruned-small/: each holds a manifest, a byte-offset index, and an individual file per chunk. .load-partial directly accesses the manifest from this repository and acquires only the chunks it is instructed to obtain; see Sharding and Partial Loading.

Exporting Derivations

.run-export <file> runs inference on a loaded network and writes what that run derives, along with any contradictions encountered, into a JSON Lines file: each line corresponds to a single derived fact or contradiction, including the premises that produced it. Every node is listed under each of its names – in a Wikidata network, this includes its Q or P identifier and its English label – enabling a converter to transform the file into any desired format, such as compact training data suitable for a language model. The structure is detailed in Exporting Derivations.

Citation

If you use this dataset, cite as:

@dataset{zelph,
  author = {Stefan Zipproth},
  title = {zelph Binaries Dataset},
  year = {2026},
  url = {https://huggingface.co/datasets/acrion/zelph}
}