farbodtavakkoli/OTel-2.0-LLM-31B-IT
OTel-2.0-LLM-31B-IT by farbodtavakkoli is a 31 billion parameter instruction-tuned language model, post-trained from Gemma 4 31B-IT on approximately 440 billion telecom training tokens. This model is specialized for telecommunications, excelling in network operations, standards interpretation, product development, and telecom-specific question answering. It utilizes Orthogonal Subspace Fine-Tuning (OSFT) to integrate new domain knowledge while preserving general instruction-following capabilities, making it ideal for RAG over telecom standards and configuration assistance.
Loading preview...
OTel-2.0-LLM-31B-IT: Telecom-Specialized LLM
OTel-2.0-LLM-31B-IT is a 31 billion parameter instruction-tuned model, developed by AT&T Chief Data Office, building upon the Gemma 4 31B-IT base. It has undergone extensive post-training on approximately 440 billion telecom-specific tokens, significantly expanding its domain knowledge compared to its predecessor, OTel 1.0. This model leverages Orthogonal Subspace Fine-Tuning (OSFT), a method that allows the model to absorb new telecom domain knowledge while largely preserving its existing general instruction-following capabilities, mitigating catastrophic forgetting.
Key Capabilities
- Telecom Domain Expertise: Specialized in telecommunications, trained on a vast corpus from 3GPP, ETSI, ITU, GSMA, CAMARA, O-RAN, and TM Forum.
- Enhanced Data Mixture: Supports RAG, abstention, direct telecom QnA, general-purpose instruction following, and agentic tool calling (though not telecom-specific tool calling).
- Multimodal Architecture (Text-Only Training): While architecturally multimodal (text + image), only the text path was trained and evaluated for telecom specialization. The vision encoder is stock Gemma 4.
- Robust Training: Processed over 1 trillion tokens using Red Hat's open-source Synthetic Data Generation Hub.
Good For
- Retrieval-Augmented Generation (RAG): Ideal for generating answers grounded in telecom standards and technical documentation.
- Standards Interpretation: Excels at summarizing and interpreting materials from major telecom standards bodies.
- Network Operations & Engineering: Supports product development, network configuration assistance, and telecom-specific direct QnA.
- Agentic Workflows: Suitable for workflows where an external system provides verified tools and validation, leveraging the model's general-purpose tool-calling training.