farbodtavakkoli/OTel-LLM-14B-IT

OTel-LLM-14B-IT: A Specialized Telecom Language Model

OTel-LLM-14B-IT is a 14 billion parameter language model developed by farbodtavakkoli, specifically fine-tuned for the telecommunications domain. It is part of the OTel Family of Models, an open-source initiative focused on creating AI resources for the global telecom sector. This model is built on the Qwen/Qwen3-14B architecture and has undergone full-parameter post-training using the proprietary OTel-LLM dataset.

Key Capabilities and Differentiators

• Domain-Specific Accuracy: Achieves a +5.5 percentage point improvement in context-grounded correctness over its base model on OTel evaluation splits, specifically for telecom-related queries.
• Context-Grounded RAG: Optimized for Retrieval-Augmented Generation (RAG) pipelines, excelling at generating answers based only on provided telecom contexts.
• Specialized Training Data: Trained on a curated dataset of 1.1 million telecom-focused training points, filtered to 326,767 high-confidence examples, contributed by institutions like Yale University, GSMA, and Khalifa University.
• Robust Training: Utilizes advanced training techniques including AdamW optimizer, cosine decay learning rate, BF16 precision, and Flash Attention 2.

Intended Use Cases

• Telecom RAG Systems: Ideal for applications requiring precise, context-grounded answers within the telecommunications industry.
• Domain-Specific QA: Suitable for question-answering systems where the model is provided with relevant telecom documents or data.

Limitations

It is crucial to note that OTel-LLM-14B-IT is domain-specific and not intended for general-purpose use or unrestricted context-free question answering. Its performance is optimized for English-only, text-centric telecom tasks, and external benchmark transfer or multilingual performance should be evaluated separately.