And new data release: in partnership with @GSMA we publish Telco-Common-Corpus, the largest fully open corpus for the telecom sector, 10 billion tokens all under free license/public domain.
No Digg Deeper questions have been answered for this story yet.
GSMA evaluation suite recently showed that even frontier models significantly lagged on telecom knowledge and tasks and this even did not progress significantly over the last two years. That's mostly because this content is not being harvested by web crawlers.
And new data release: in partnership with @GSMA we publish Telco-Common-Corpus, the largest fully open corpus for the telecom sector, 10 billion tokens all under free license/public domain.
*Announcement on GSMA: https://www.open-telco.ai/telco-common-corpus-an-open-verified-data-commons-for-telecom-ai/ *Our blog: https://pleias.ai/blog/telco%20 *Joint release on @huggingface: https://huggingface.co/datasets/GSMA/Telco-Common-Corpus and https://huggingface.co/datasets/PleIAs/Telco-Common-Corpus
Now the release is just an initial step: having the content freely shareable means it can propel data research at scale, especially toward generated environment for search, code, infrastructure simulation. More to come.
The actual discovery that validates our open data paradox: most of this content is actually shareable. Industrial information is widely available in patents, scientific publications, wikipedia article, preparatory documents from standard bodies.
GSMA evaluation suite recently showed that even frontier models significantly lagged on telecom knowledge and tasks and this even did not progress significantly over the last two years. That's mostly because this content is not being harvested by web crawlers.
Yet, the original raw sources were not really usable for training: we had to transform millions of PDFs and docx documents into structured markdown with preserved artifacts (tables, footnotes) using our new text processing infrastructure, Stratum.
The actual discovery that validates our open data paradox: most of this content is actually shareable. Industrial information is widely available in patents, scientific publications, wikipedia article, preparatory documents from standard bodies.
Now the release is just an initial step: having the content freely shareable means it can propel data research at scale, especially toward generated environment for search, code, infrastructure simulation. More to come.
Now the release is just an initial step: having the content freely shareable means it can propel synthetic data research at scale and generated environment for search, code, infrastructure simulation. More to come.
Yet, this content is not directly usable for training. This is our first release powered by new text processing infrastructure, Stratum. Beyond PDfs, we had millions of docx document and it proved more effective to emulate and re-digitize them using our internal VLMs.