This paper proposes TelecomGPT, a framework to build telecom-specific large language models (LLMs). The authors address the challenge of adapting general-purpose LLMs to telecom-specific tasks by collecting and building three specialized datasets: a pre-training dataset, an instruction dataset, and a preference dataset. These datasets are used for continual pre-training, instruction tuning, and alignment tuning, respectively. The authors also propose three new evaluation benchmarks: Telecom Math Modeling, Telecom Open QnA, and Telecom Code Tasks, which comprehensively assess the capabilities of LLMs in telecom-related tasks. The fine-tuned LLM TelecomGPT outperforms state-of-the-art LLMs such as GPT-4, Llama-3, and Mistral in the Telecom Math Modeling benchmark and achieves comparable performance in other benchmarks. The authors also discuss the challenges of building telecom-specific LLMs, including the lack of open-source datasets, the dominance of technical documents, multi-modality contents, and frequent knowledge updates. The methodology involves three stages: domain-specific continual pre-training, instruction tuning, and alignment tuning. The authors also describe the dataset construction process, including the collection of telecom standards, research papers, and patents, and the filtering of telecom-relevant content from general datasets. The instruction dataset includes various telecom-related tasks such as multiple-choice question answering, open-ended question answering, technical documents classification, math modeling, code generation, code infilling, code summary, and code analysis. The alignment dataset is used to align the model with human preferences. The authors evaluate the performance of TelecomGPT on various benchmarks, including Telecom MCQ, Telecom Standard Classification, Telecom Math Modeling, Telecom Code Understanding and Generation, Telecom Instruct Following, and others. The results show that TelecomGPT outperforms other LLMs in several benchmarks, particularly in Telecom Math Modeling and Telecom Code Understanding and Generation. The authors conclude that TelecomGPT is a promising approach for building telecom-specific LLMs that can effectively handle telecom-related tasks.This paper proposes TelecomGPT, a framework to build telecom-specific large language models (LLMs). The authors address the challenge of adapting general-purpose LLMs to telecom-specific tasks by collecting and building three specialized datasets: a pre-training dataset, an instruction dataset, and a preference dataset. These datasets are used for continual pre-training, instruction tuning, and alignment tuning, respectively. The authors also propose three new evaluation benchmarks: Telecom Math Modeling, Telecom Open QnA, and Telecom Code Tasks, which comprehensively assess the capabilities of LLMs in telecom-related tasks. The fine-tuned LLM TelecomGPT outperforms state-of-the-art LLMs such as GPT-4, Llama-3, and Mistral in the Telecom Math Modeling benchmark and achieves comparable performance in other benchmarks. The authors also discuss the challenges of building telecom-specific LLMs, including the lack of open-source datasets, the dominance of technical documents, multi-modality contents, and frequent knowledge updates. The methodology involves three stages: domain-specific continual pre-training, instruction tuning, and alignment tuning. The authors also describe the dataset construction process, including the collection of telecom standards, research papers, and patents, and the filtering of telecom-relevant content from general datasets. The instruction dataset includes various telecom-related tasks such as multiple-choice question answering, open-ended question answering, technical documents classification, math modeling, code generation, code infilling, code summary, and code analysis. The alignment dataset is used to align the model with human preferences. The authors evaluate the performance of TelecomGPT on various benchmarks, including Telecom MCQ, Telecom Standard Classification, Telecom Math Modeling, Telecom Code Understanding and Generation, Telecom Instruct Following, and others. The results show that TelecomGPT outperforms other LLMs in several benchmarks, particularly in Telecom Math Modeling and Telecom Code Understanding and Generation. The authors conclude that TelecomGPT is a promising approach for building telecom-specific LLMs that can effectively handle telecom-related tasks.