ChartGemma is a novel chart understanding and reasoning model developed over PaliGemma, designed to address the limitations of existing chart representation models. Unlike previous methods that rely on underlying data tables for instruction-tuning, ChartGemma is trained on instruction-tuning data directly generated from chart images, capturing both high-level trends and low-level visual information. This approach enables the model to learn from a diverse set of charts and better understand complex visual patterns. ChartGemma achieves state-of-the-art results across five benchmarks, including chart summarization, question answering, and fact-checking, and produces more realistic and factually correct summaries compared to its contemporaries. The model is also significantly smaller than existing chart understanding models, making it suitable for real-world applications. ChartGemma's architecture features a strong backbone model and more representative instruction-tuning data, allowing it to effectively tackle existing benchmarks. The model's instruction-tuning data is generated directly from chart images, enabling it to capture complex visual elements and improve its ability to understand and represent real-world charts. ChartGemma's performance is evaluated on various benchmarks, including ChartQA, ChartFC, and ChartCheck, where it outperforms or matches existing models. The model also performs well on open-ended tasks, such as chart-to-text summarization and open-ended question answering. Human evaluations show that ChartGemma produces more informative and factually correct outputs compared to other models. Despite its effectiveness, ChartGemma has some limitations, including the use of a proprietary LLM for instruction-tuning data generation, a fixed input resolution for the vision encoder, and reliance on a closed-source model for evaluating key metrics. The model is also prone to hallucinations and may produce factually incorrect statements or erroneous code. Overall, ChartGemma is an effective model for understanding and reasoning over real-world charts, with a strong performance on various benchmarks and a smaller size compared to existing models.ChartGemma is a novel chart understanding and reasoning model developed over PaliGemma, designed to address the limitations of existing chart representation models. Unlike previous methods that rely on underlying data tables for instruction-tuning, ChartGemma is trained on instruction-tuning data directly generated from chart images, capturing both high-level trends and low-level visual information. This approach enables the model to learn from a diverse set of charts and better understand complex visual patterns. ChartGemma achieves state-of-the-art results across five benchmarks, including chart summarization, question answering, and fact-checking, and produces more realistic and factually correct summaries compared to its contemporaries. The model is also significantly smaller than existing chart understanding models, making it suitable for real-world applications. ChartGemma's architecture features a strong backbone model and more representative instruction-tuning data, allowing it to effectively tackle existing benchmarks. The model's instruction-tuning data is generated directly from chart images, enabling it to capture complex visual elements and improve its ability to understand and represent real-world charts. ChartGemma's performance is evaluated on various benchmarks, including ChartQA, ChartFC, and ChartCheck, where it outperforms or matches existing models. The model also performs well on open-ended tasks, such as chart-to-text summarization and open-ended question answering. Human evaluations show that ChartGemma produces more informative and factually correct outputs compared to other models. Despite its effectiveness, ChartGemma has some limitations, including the use of a proprietary LLM for instruction-tuning data generation, a fixed input resolution for the vision encoder, and reliance on a closed-source model for evaluating key metrics. The model is also prone to hallucinations and may produce factually incorrect statements or erroneous code. Overall, ChartGemma is an effective model for understanding and reasoning over real-world charts, with a strong performance on various benchmarks and a smaller size compared to existing models.