StructLM is a generalist model designed to handle structured knowledge grounding (SKG) tasks. The paper introduces a large instruction-tuning dataset with 1.1 million examples, used to train a series of models based on Mistral and CodeLlama, ranging from 7B to 34B parameters. StructLM outperforms task-specific models on 16 out of 18 evaluated datasets and achieves state-of-the-art performance on 8 SKG tasks. It also shows strong generalization across 6 novel held-out tasks, outperforming TableLlama by 35% and Flan-UL2 20B by 10%. Despite the model size increasing from 7B to 34B, the performance gains are marginal, suggesting that SKG remains a challenging task requiring more innovative design. The model weights and training dataset are released to the community. StructLM demonstrates strong zero-shot generalization capabilities on unseen SKG tasks, surpassing previous models. The paper also explores the effects of different pretraining data and instruction-tuning approaches on model performance, finding that code pretraining is most effective. StructLM's performance is evaluated on various SKG tasks, including data-to-text generation, table-based question answering, knowledge-grounded conversations, fact verification, SQL generation, and mathematical reasoning. The results show that StructLM significantly outperforms existing models on multiple SKG tasks, highlighting its effectiveness in handling structured data. The paper also discusses the importance of structured data in language models and the potential for future research in this area.StructLM is a generalist model designed to handle structured knowledge grounding (SKG) tasks. The paper introduces a large instruction-tuning dataset with 1.1 million examples, used to train a series of models based on Mistral and CodeLlama, ranging from 7B to 34B parameters. StructLM outperforms task-specific models on 16 out of 18 evaluated datasets and achieves state-of-the-art performance on 8 SKG tasks. It also shows strong generalization across 6 novel held-out tasks, outperforming TableLlama by 35% and Flan-UL2 20B by 10%. Despite the model size increasing from 7B to 34B, the performance gains are marginal, suggesting that SKG remains a challenging task requiring more innovative design. The model weights and training dataset are released to the community. StructLM demonstrates strong zero-shot generalization capabilities on unseen SKG tasks, surpassing previous models. The paper also explores the effects of different pretraining data and instruction-tuning approaches on model performance, finding that code pretraining is most effective. StructLM's performance is evaluated on various SKG tasks, including data-to-text generation, table-based question answering, knowledge-grounded conversations, fact verification, SQL generation, and mathematical reasoning. The results show that StructLM significantly outperforms existing models on multiple SKG tasks, highlighting its effectiveness in handling structured data. The paper also discusses the importance of structured data in language models and the potential for future research in this area.