Red Hat details the open training pipeline behind telecom model OTel 2.0

Red Hat has published the engineering path behind OTel 2.0, a domain model trained for telecommunications work: collect standards documents, generate several forms of synthetic instruction data, fine-tune every model weight, then add domain-aware safety testing before deployment.

The Aug. 26 account gives unusually concrete scale figures. Red Hat says GSMA contributed a corpus of about 15 billion raw tokens drawn from seven standards bodies. AT&T then used Red Hat’s open-source SDG Hub to process more than one trillion tokens and produce roughly 440 billion training tokens for OTel 2.0.

Turning standards into instruction data

The source material spans specifications and documents from organizations including 3GPP, ETSI, GSMA, CAMARA, ITU, O-RAN and TM Forum. Rather than apply one synthetic-data recipe, the SDG Hub workflow combines four: direct question-and-answer generation from complete sections, extractive summaries, detailed thematic summaries and atomic key facts.

That design is meant to teach both terminology and relationships. A model needs more than the definition of a radio procedure; it must connect procedural steps to the wider network context in which an operator would diagnose a fault or choose an action.

Red Hat says AT&T performed the data-preparation work on Microsoft Managed Compute with about 530 GPUs, primarily AMD MI300X accelerators. AT&T then used supervised fine-tuning with full-weight updates on on-premises AMD MI355X systems supplied through Dell infrastructure. That is a materially heavier training path than parameter-efficient adaptation, but it allows the model to absorb domain terminology across all parameters.

The open components extend beyond training

The article positions OTel 2.0 inside a wider Linux Foundation Networking observability proof of concept. In that design, an inference engine uses the model to interpret network telemetry, relate faults to 3GPP standards and generate explanations for a closed-loop automation flow. Red Hat serves as technical lead for the working group developing the guide and simulated proof of concept.

Red Hat also describes a safety pipeline built from SDG Hub-generated adversarial cases, the Garak vulnerability scanner and production guardrails. The proposed tests include prompt injection, fabricated network parameters and attempts to bypass constraints. Those controls are especially relevant when model output can feed operational automation rather than a read-only assistant.

A reusable pattern, with validation still required

The important deliverable is broader than a telecom model. SDG Hub’s extraction flows can be applied to any authoritative technical corpus, and Red Hat’s Training Hub exposes Orthogonal Subspace Fine-Tuning as one possible future approach to continual learning without overwriting earlier capabilities.

But an open pipeline does not remove the need for evidence. Red Hat reports more than five million OTel 2.0 downloads and plans weekly model-weight updates, while the public engineering account does not present task-level accuracy or safety results. Teams evaluating the model should therefore treat the training recipe, benchmarks and deployment controls as three separate artifacts.

The project nevertheless offers a concrete blueprint for domain AI: standards-owned source data, reproducible generation and training components, explicit infrastructure, and safety checks tied to the operating environment.