SaulLM-7B is a large language model (LLM) specifically designed for the legal domain, with 7 billion parameters. It is trained on an English legal corpus of over 30 billion tokens, leveraging the Mistral 7B architecture. The model is enhanced through instruction fine-tuning using legal datasets, improving its performance in legal tasks. SaulLM-7B is released under the MIT License and is part of a family of legal LLMs tailored for legal text comprehension and generation. The model's training involves extensive pretraining on legal corpora from English-speaking jurisdictions, along with instruction fine-tuning to enhance legal task performance. The paper introduces LegalBench-Instruct, an improved evaluation protocol for legal LLMs, and includes legal tasks from the MMLU benchmark. The model, SaulLM-7B, and its instruction-tuned variant, SaulLM-7B-Instruct, are made available for research and development. The model demonstrates superior performance in legal tasks compared to other models, with significant improvements in legal understanding and application. The paper also discusses the model's performance on various legal benchmarks and its potential to advance legal language understanding and application. The model is trained on a diverse legal corpus, including data from the U.S., Europe, and Australia, and is evaluated on legal documents, legal decisions, legislation, and party submissions. The model's performance is measured using perplexity and balanced accuracy, showing significant improvements over existing models. The paper concludes that SaulLM-7B represents a significant step forward in the development of legal LLMs, with potential applications in legal research and practice.SaulLM-7B is a large language model (LLM) specifically designed for the legal domain, with 7 billion parameters. It is trained on an English legal corpus of over 30 billion tokens, leveraging the Mistral 7B architecture. The model is enhanced through instruction fine-tuning using legal datasets, improving its performance in legal tasks. SaulLM-7B is released under the MIT License and is part of a family of legal LLMs tailored for legal text comprehension and generation. The model's training involves extensive pretraining on legal corpora from English-speaking jurisdictions, along with instruction fine-tuning to enhance legal task performance. The paper introduces LegalBench-Instruct, an improved evaluation protocol for legal LLMs, and includes legal tasks from the MMLU benchmark. The model, SaulLM-7B, and its instruction-tuned variant, SaulLM-7B-Instruct, are made available for research and development. The model demonstrates superior performance in legal tasks compared to other models, with significant improvements in legal understanding and application. The paper also discusses the model's performance on various legal benchmarks and its potential to advance legal language understanding and application. The model is trained on a diverse legal corpus, including data from the U.S., Europe, and Australia, and is evaluated on legal documents, legal decisions, legislation, and party submissions. The model's performance is measured using perplexity and balanced accuracy, showing significant improvements over existing models. The paper concludes that SaulLM-7B represents a significant step forward in the development of legal LLMs, with potential applications in legal research and practice.