---
license: cc-by-4.0
task_categories:
- graph-ml
- text-retrieval
language:
- en
tags:
- knowledge-graph
- telecommunications
- 3gpp
- citation-network
- graph-rag
pretty_name: 3GPP Spec Citation Graph
size_categories:
- 100K<n<1M
---

# 3GPP Spec Citation Graph

A clause-level cross-reference graph over 3GPP technical specifications,
covering **Releases 11, 12, 13, 14, and 16** (Release 15 and releases outside
this band are not included — see Coverage below). Built from the
[TSpec-LLM](https://huggingface.co/datasets/rasoul-nikbakht/TSpec-LLM) corpus.
Nodes are individual specification **clauses** (not whole documents); edges
are explicit cross-references extracted from spec text, typed by reference
granularity (document-level citation, intra-document clause pointer, etc.).

Released standalone, independent of any particular retrieval pipeline, in
the spirit of
[joyboseroy/darshana-graph](https://huggingface.co/datasets/joyboseroy/darshana-graph):
the graph structure itself is the artifact.

**This is a citation/cross-reference graph, not a general-purpose 3GPP
knowledge graph.** Several entity-relation KGs already exist for 3GPP
content (e.g. `otellm/3gpp_knowledgeGraph` for Release 13, `GSMA/telecom-kg-rel19`)
— those extract concepts/parameters/procedures as nodes. This dataset is
different: nodes are clauses, edges are references between them, and the
intended use is studying reference/citation topology, not entity QA.

## Dataset Structure

**`nodes.parquet`** — 130,270 rows
| column | type | description |
|---|---|---|
| node_id | int | internal graph node id |
| doc_id | string | source document identifier (e.g. `25173-e00`) |
| clause_id | string | clause number within the document (e.g. `4.2.2`) |
| label | string | graph node label (`Clause`) |
| release | string | inferred release tag (Rel-11/12/13/14/16), or `unknown` |

**`edges.parquet`** — 209,934 rows
| column | type | description |
|---|---|---|
| source_id / target_id | int | node ids |
| relation | string | edge type (`REFERENCES`) |
| ref_type | string | reference granularity: `document_level` (152,656), `intradoc` (53,199), `clause` (3,474), `annex` (357), `bare_lowconf` (248) |
| source_doc / source_clause / target_doc / target_clause | string | convenience fields for joins without re-merging nodes.parquet |

**`release_summary.csv`** — per-release node and outgoing-edge counts:

| release | node_count | outgoing_edge_count |
|---|---|---|
| Rel-11 | 8,852 | 14,544 |
| Rel-12 | 21,793 | 35,293 |
| Rel-13 | 25,773 | 42,537 |
| Rel-14 | 26,904 | 43,828 |
| Rel-16 | 42,969 | 73,087 |
| unknown | 3,979 (3.1%) | 645 |

## Construction

Extracted from 3GPP specification markdown files (TSpec-LLM subset covering
Rel-11–14 and Rel-16), with cross-references parsed and loaded into FalkorDB
as a property graph. `ref_type` distinguishes document-level citations from
finer-grained intra-document/clause/annex references, based on the parser's
own classification at extraction time.

## Intended Uses

- Cross-reference/citation topology analysis across 3GPP releases
- Graph-RAG retrieval benchmarking (multi-hop reference traversal)
- Network-science studies of technical-standard document ecosystems at
  clause granularity

## Limitations

- **Coverage gap**: Releases 11–14 and 16 only. Release 15 and releases
  outside this band (8–10, 17+) are not represented. Do not use this dataset
  for claims about 3GPP standards evolution as a whole without accounting
  for this gap.
- **Extraction precision**: cross-reference parsing is regex/rule-based; an
  LLM-judge spot-check (Groq llama-3.3-70b) found ~0.71 precision, meaning
  roughly 3 in 10 extracted edges may be spurious or mis-parsed. The
  `bare_lowconf` ref_type (248 edges) flags some of this at the source, but
  most noise is not self-flagged and sits inside the other ref_type categories.
- **Release tagging**: inferred from source directory structure; 3.1% of
  nodes (3,979) have `release = unknown`.
- Single edge type (`REFERENCES`) with a `ref_type` qualifier — no other
  relation types (e.g. no semantic/entity relations, only citations).

## Citation

No accompanying paper. If you use this dataset, please cite it directly:

```bibtex
@misc{bose2026_3gpp_citation_graph,
  author = {Bose, Joy},
  title = {3GPP Spec Citation Graph},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/joyboseroy/3gpp-citation-graph}}
}
```