---
title: README
emoji: 📡
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
---

_Welcome to the official GSMA organization on Hugging Face!_

The [GSMA](https://www.gsma.com) represents mobile operators and organisations across the mobile ecosystem worldwide. We are building **open resources to advance AI in telecommunications** — making telecom-domain evaluation, benchmarking, and knowledge accessible to the global research community.

## Open Telco AI

[**Open Telco**](https://github.com/gsma-research/open_telco) is a comprehensive suite of telco-specific benchmarks built on the [Inspect AI](https://inspect.ai-safety-institute.org.uk/) framework, designed to ensure safe and optimal deployment of AI in telecommunications environments. A collaborative effort with major telecom providers, research institutions, and universities.

- **[ot-full](https://huggingface.co/datasets/GSMA/ot-full)** — 16,866 evaluation samples across 7 benchmarks — the complete evaluation suite
- **[ot-lite](https://huggingface.co/datasets/GSMA/ot-lite)** — 1,700 sample subset for fast iteration during model development
- **[Leaderboard Scores](https://huggingface.co/datasets/GSMA/leaderboard)** — Published benchmark scores with standard errors

### Benchmarks

The evaluation suite curates 7 telecom-domain benchmarks from academic and industry sources:

| Benchmark | Samples | Task |
|-----------|---------|------|
| **TeleQnA** | 10,000 | Multiple-choice Q&A on telecom standards |
| **TeleMath** | 1,500 | Mathematical reasoning in telecom contexts |
| **TeleTables** | 500 | Table interpretation from 3GPP specifications |
| **TeleLogs** | 586 | Log analysis and network troubleshooting |
| **3GPP TSG** | 3,780 | 3GPP Technical Specification Group document understanding |
| **ORANBench** | 200 | O-RAN architecture and specifications |
| **SRSRANBench** | 300 | srsRAN open-source network stack |

### OTel — Open Telco AI Models

A collaborative effort to build AI models for the global telecommunications sector, optimized for RAG and agentic applications.

**Model Suite:**
- **18 Language Models** (270M–32B parameters)
- **10 Embedding Models** (22M–8B parameters)
- **3 Reranker Models** (0.6B–8B parameters)

Models are trained on telecom-domain data — including 3GPP specifications, O-RAN documentation, and RFC standards — curated by 200+ domain experts from AT&T, GSMA, Purdue University, Khalifa University, University of Leeds, Yale University, and others.

**Resources:**
- [Training & Inference Code (GitHub)](https://github.com/farbodtavakkoli/OTel)
- [Open-Telco-1 Dataset](https://huggingface.co/datasets/GSMA/Open-Telco-1)
- [LLM Collection](https://huggingface.co/collections/farbodtavakkoli/otel-llm)
- [Embedding Collection](https://huggingface.co/collections/farbodtavakkoli/otel-embedding)
- [Reranker Collection](https://huggingface.co/collections/farbodtavakkoli/otel-reranker)

**License:** Apache-2.0

### Additional Resources

## Models
- **[AdaptKey-Nemotron-30b](https://huggingface.co/AdaptKey/AdaptKey-Nemotron-30b)** — NVIDIA Nemotron 3 Nano fine-tuned by AdaptKey for telecom. *Contributed by NVIDIA and AdaptKey.*

## Datasets
- **[telecom-kg-rel19](https://huggingface.co/datasets/GSMA/telecom-kg-rel19)** — Large-scale telecom knowledge graph built from 3GPP Release 19 specifications, with text chunks for retrieval-augmented generation (RAG) and LLM reasoning over standards
- **[oran_spec_knowledge_graph](https://huggingface.co/datasets/GSMA/oran_spec_knowledge_graph)** A knowelge graph of 25,103 nodes and 98,679 relationships extracted from official O-RAN Alliance specification documents using OpenAI GPT-4.1
- **[AdaptKey Nemotron 30B Training Data](https://huggingface.co/AdaptKey/AdaptKey-Nemotron-30b/tree/main/training_data)** — Dataset used to fine-tune Nemotron 3 Nano for telecom. *Contributed by NVIDIA and AdaptKey.*

## Guides & Blueprints

- **[NVIDIA Blueprint: AI Agent for Telecom Network Configuration Planning](https://build.nvidia.com/nvidia/telco-network-configuration)** — Agentic blueprint for RAN configuration. *Contributed by NVIDIA (in collaboration with BubbleRAN).*
- **[NVIDIA Blueprint: Intent Driven RAN Energy Efficiency](https://build.nvidia.com/viavi/intent-driven-ran-energy-efficiency)** — Agentic blueprint for RAN energy saving with simulation. *Contributed by NVIDIA (in collaboration with VIAVI).*
- **[NVIDIA Guide: Teaching a Model to Reason Over Telecom Network Incidents](https://nvidia-nemo.github.io/Skills/tutorials/2026/02/27/teaching-a-model-to-reason-over-telecom-network-incidents/)** — Guide on how to build NOC reasoning agents. *Contributed by NVIDIA (in collaboration with Tech Mahindra).*

## Satellite — Eval Runner

[**Satellite**](https://github.com/gsma-labs/satellite) provides telecom-focused evaluation operations built on Inspect AI. Run the full Open Telco benchmark suite locally within your own infrastructure with a single command.

## Telecom Simulation Sandboxes

Purpose-built sandbox environments that place AI agents inside live telecom network simulations — for evaluating whether models can *operate* networks, not just answer questions about them.

- **[inspect-kathara](https://github.com/gsma-labs/inspect-kathara)** — Run AI agent evaluations inside isolated network topologies. Integrates Inspect AI with Docker-based network sandboxes to evaluate agents' ability to diagnose and resolve network connectivity issues in reproducible environments.

- **[5gs-sandbox](https://github.com/gsma-labs/5gs-sandbox)** — Run AI agent evaluations inside a complete 5G Standalone network. A full 5G SA deployment with 15 Docker containers (Open5GS + UERANSIM), enabling agents to configure, diagnose, and optimize real 5G network functions with actual performance measurement.

## Research and Community

- **GSMA AI Initiatives**: https://www.gsma.com/solutions-and-impact/technologies/artificial-intelligence/open-telco.ai/
- **Open Gateway**: https://www.gsma.com/solutions-and-impact/gsma-open-gateway/
- **MWC (Mobile World Congress)**: https://www.mwcbarcelona.com/