Telco-Common-Corpus / README.md
Pclanglais's picture
Update README.md
c590e4e verified
|
Raw
History Blame Contribute Delete
2.67 kB
metadata
language:
  - en

Telco Common Corpus (TCC) is a ten billion tokens collection of fully open, free licensed telecommunications knowledge (scientific literature, patents, open data, and open-web projects) with licence and provenance verified at a document-level.

TCC stems from GSMA's effort to make AI work for the telecom sector. The Open-Telco LLM Benchmarks and the broader Open Telco AI initiative have already established that current models fall short on real telecom tasks, including network management and, even more importantly, that they have not progressed significantly over the last two years.

Composition

What equally sets TCC apart is the diversity of its sources: alongside peer-reviewed articles, it reaches technical reports, standards-adjacent project deliverables, patents, and public-domain government research on propagation, spectrum, and coding.

Source What it covers Licence Tokens
3GPP preparatory documents Working documents behind the 3GPP standards Public domain 3,546,022,615
RFC specifications ~10,000 IETF specifications Public domain 113,005,444
RFC drafts ~40,000 updated drafts Public domain 307,750,151
RFC preparatory documents Spec drafts, discussions, and related artifacts Public domain / no copyright 1,374,532,228
RFC proceedings Minutes and discussions Public domain / no copyright 47,930,212
IEEE open access 64,000 journal articles released by IEEE under free licence Various, mostly CC-BY 1,210,317,934
Other open access papers from OpenAlex Various, mostly CC-By 459,392,269
US patents 95,000 US patents (2020–2026), filtered to the telecom domain Public domain 2,345,121,436
EU patents 25,000 EU patents (last two years), filtered to the telecom domain Public domain 606,815,778
Wikipedia (telecom) Structured Wikipedia filtered to telecom topics CC-BY-SA 12,132,156
Wikidata (telecom) Structured Wikipedia filtered to telecom topics CC0 11,345,452
Total 10,034,365,675

Processing

Most TCC sources come from doc or PDF documents that require advanced processing. We used a Pleias internal pipeline (Stratum) relying on open weights VLMs like dots.ocr to properly parse each document at the section level and preserve fundamental document structure like tables.

The dataset is almost exclusively English. The few multilingual documents have been detected using CommonLingua, a SOTA language detection already jointly released by Pleias and GSMA.