Distro News

Open Telco AI uses SDG Hub to train OTel 2.5 for autonomous network operations

Key Takeaway

Red Hat and AT&T are using synthetic data generation to teach the OTel 2.5 model how to use Model Context Protocol (MCP) tools for telecom network management.

Summary

The Open Telco AI initiative has announced advancements in training the OTel 2.5 model to perform operational tasks within telecommunications networks. By utilizing Red Hat's open-source Synthetic Data Generation HUB (SDG Hub), the project converts complex industry standards from organizations like 3GPP and TM Forum into actionable training data. This moves AI capabilities beyond mere specification recitation toward autonomous execution.

The training process focuses on teaching the model to interact with network tools via the Model Context Protocol (MCP). This allows the AI to query active alarms, verify router states, and manage trouble tickets. The workflow leverages Red Hat OpenShift and OpenShift AI to generate validated tool-calling datasets, aiming to move service providers toward Level 4 autonomous network operations.

Technical implementation of OTel 2.5 training

  • Tool Integration: The system utilizes eight MCP servers providing 84 distinct tools based on TM Forum Open APIs (TMF621, TMF628, TMF642), CAMARA APIs, and 3GPP specifications (NWDAF and UDR).
  • Synthetic Data Pipeline: SDG Hub employs an MCP distillation flow where a teacher model explores tool schemas, generates realistic engineering questions, records tool-use demonstrations, and filters results based on quality thresholds.
  • Execution Validation: To prevent hallucinations, the pipeline enforces mandatory tool verification, validates all tool responses against OpenAPI schemas, and isolates timeout-related data to ensure dataset integrity.
  • Training Complexity: The pipeline organizes tasks into five progressive levels: single-tool tasks, parallel calls, sequential chains, multi-turn conversations, and closed-loop workflows connecting alarms to trouble tickets.
  • Infrastructure: The workflow runs as a Kubeflow pipeline on Red Hat OpenShift AI, producing JSONL training conversations for model fine-tuning.

Why it matters

This development provides a blueprint for using synthetic data to train domain-specific AI agents in highly regulated industries where real-world troubleshooting data is scarce. For telecom engineers and cloud-native developers, it demonstrates how the Model Context Protocol (MCP) can bridge the gap between large language models and critical network infrastructure.

Original reference: www.redhat.com →