--- license: cc-by-sa-4.0 language: - en size_categories: - 1M:` | | `subject_label` | canonical English label | | `subject_description` | English description | | `subject_instance_of_ids` | pipe-separated P31 Q-IDs | | `subject_instance_of_labels` | pipe-separated P31 labels | | `property_id` | P-ID (`P31`, `P50`, `P577`, `P1476`, …) | | `property_label` | property label | | `value_id` | Q-ID or string value | | `value_label` | value label | | `value_type` | `wikibase-entityid` / `string` / `time` / `monolingualtext` / `url` | | `qualifier_property_id` | qualifier P-ID (may be empty) | | `qualifier_property_label` | qualifier label | | `qualifier_value_id` | qualifier value | | `qualifier_value_label` | qualifier value label | A statement with N qualifiers becomes N rows sharing the same `statement_id`; a statement with no qualifier is one row with empty qualifier fields. ## Sources | source | distinct subjects | rows | |---|---:|---:| | `wikidata` | 287,535 | ~4.55M | | `IETF` (RFCs) | 9,757 | 104,284 | | `3GPP` (TS/TR) | 3,596 | 25,644 | | `ITU-T` (recommendations) | 738 | 2,952 | | `Bluetooth` (SIG specs) | 257 | 1,028 | | `CableLabs` (DOCSIS) | 66 | 264 | ## Wikidata side All wikidata properties of every kept subject are preserved — **no property selection**. The pipeline filters out *entities* (e.g. radio-personality humans, ships, off-topic content stubs) but never drops individual statements from a kept entity. Build summary: 75 hand-curated telecom anchor classes → P279 BFS walk → 6,448 classes → P31 entity scan → property-based expansion (P452 / P101 / P106 / P921 / P136 / P1056) → anchor cascade (P108 / P127 / P137 / P749 / P176 / P361 / P488 / P169 / P3320 / P710) → concept-anchor patch (MIMO, OFDM, RSRP, IEEE 802.11n/ac/ax, Shannon-Hartley, AWGN, Friis, …) → weak-area patch (pioneers, ITU-T codecs, country regulators, SDN/NFV, mobile OS) → cleanup filters → **287,535 wikidata subjects**. ## Standards side | SDO | Source | Notes | |---|---|---| | IETF | `rfc-editor.org/rfc-index.xml` | full metadata: title, authors, year/month, status, abstract, obsoletes/-by, keywords | | 3GPP | `3gpp.org/ftp/Specs/archive/` + `3gpp.org/dynareport` | 91 % titles + last_modified + per-version date | | ITU-T | `itu.int/itu-t/recommendations` per series | std_id + series | | Bluetooth SIG | `bluetooth.com/specifications/specs/` | title + url | | CableLabs / DOCSIS | `cablelabs.com/specifications` | title + url | Standards rows use a **13-property wikidata-aligned subset** (P31, P1476 title, P577 date, P361 part-of, P953 URL, P50 author, P155/P156 follows/-by, P393 edition, P348 version, P407 language, P5008 status, P1683 abstract) — a deliberate schema alignment to make standards joinable against the wikidata rows. Native per-SDO catalog parquets with all original fields are preserved upstream. Not ingested (JS-rendered SDO catalogs): ETSI, GSMA, IEEE Get Program. ## Subject-ID convention - `Q…` — wikidata entity - `STD:IETF:RFC0791`, `STD:3GPP:TS_38.211`, `STD:ITU-T:G.711`, `STD:Bluetooth:…`, `STD:CableLabs:…` — standards Synthetic instance-of classes for SDO docs without wikidata Q-ID counterparts: `SYN:3GPP-TS`, `SYN:3GPP-TR`, `SYN:BT-SPEC`, `SYN:DOCSIS`. ## Loading ```python import pyarrow.parquet as pq, pyarrow.compute as pc # stream pf = pq.ParquetFile("telecom_kb_v15.parquet") for batch in pf.iter_batches(batch_size=50_000): for row in batch.to_pylist(): pass # filter examples t = pq.read_table("telecom_kb_v15.parquet") verizon = t.filter(pc.equal(t["subject_id"], "Q467752")) ietf = t.filter(pc.equal(t["source"], "IETF")) ``` Reconstruct a per-entity view by grouping on `subject_id` (+ `statement_id`). ## Caveats - Wikidata side reflects a 2026-Q1 dump; later additions are absent. - Standards side is **metadata only** (titles / dates / versions / URLs / authors / obsoletes-chains), not full spec text. - Wikidata coverage of individual standards documents is sparse: 0 3GPP TS docs by exact label, ~95 RFCs by exact `RFC NNN` label (plus ~1,700 by content-title via Q212971 RFC class), ~572 ITU-T recommendations. The standards-side rows are largely net-new vs wikidata. - Some cascade noise remains (non-telecom patents owned by Samsung / Qualcomm etc.) — documented in audits but kept per scope.