farbodtavakkoli/OTel-2.0-LLM-31B-IT
OTel-2.0-LLM-31B-IT is a 31 billion parameter instruction-tuned language model developed by farbodtavakkoli, post-trained from Gemma 4 31B-IT on approximately 440 billion telecom training tokens. This model specializes in telecommunications-specific AI workflows, including network operations, standards interpretation, product development, network configuration assistance, RAG, and telecom-specific question answering. It utilizes Orthogonal Subspace Fine-Tuning (OSFT) to integrate new domain knowledge while preserving the base model's general instruction-following capabilities, supporting a 32768 token context length.
Loading preview...
OTel-2.0-LLM-31B-IT: Telecom-Specialized Language Model
OTel-2.0-LLM-31B-IT is a 31 billion parameter instruction-tuned model, post-trained from Google's Gemma 4 31B-IT by farbodtavakkoli (AT&T Chief Data Office). It is specifically designed for telecommunications AI workflows, having been trained on approximately 440 billion telecom-specific tokens. This model represents the first release in the OTel 2.0 family, significantly expanding upon the original OTel effort with a broader and deeper telecom corpus.
Key Capabilities
- Telecom Domain Expertise: Specialized in network operations, standards interpretation (3GPP, ETSI, ITU, GSMA, CAMARA, O-RAN, TM Forum), product development, and network configuration assistance.
- Enhanced RAG and QnA: Designed for Retrieval-Augmented Generation (RAG) over telecom standards and direct telecom-specific question answering.
- Continual Learning with OSFT: Utilizes Orthogonal Subspace Fine-Tuning (OSFT) from Red Hat AI Innovation Team, allowing it to absorb new telecom domain knowledge without significantly degrading the base model's general instruction-following abilities.
- Broad Data Mixture: Incorporates direct telecom QnA, abstention, RAG, base-model-style telecom data, and general-purpose instruction-following and tool-calling examples.
- Multimodal Architecture (Text-Only Training): While architecturally multimodal and capable of accepting image inputs, its training and quality claims are text-only. The vision encoder is stock Gemma 4 and has not been domain-tuned for telecom images.
Good For
- Retrieval-Augmented Generation: Ideal for RAG applications over telecom standards, specifications, and internal technical documentation.
- Standards Interpretation: Summarizing and interpreting materials from major telecom standards bodies.
- Network Engineering Support: Assisting with product development, network configuration, and engineering tasks in a telecom context.
- Telecom-Specific QnA: Answering factual and knowledge-based questions related to telecom standards, protocols, and network concepts.
- Agentic Workflows: Suitable for integration into agentic systems where external tools, validation, and human review are provided, though it lacks telecommunications-specific tool-calling examples in its current training mixture.