This page provides prerequisites and detailed instructions for fine-tuningGemini on document data using supervised learning.

Use cases

Fine-tuning lets you customize powerful language models for your specific needs.Here are some key use cases where fine-tuning with your own set of PDFs cansignificantly enhance a model's performance:

• Internal knowledge base: Convert your internal documents into an AI-powered knowledge base that provides instant answers and insights. For example, a sales representative could instantly access product specifications and pricing details from past training materials.
• Research assistant: Create a research assistant capable of analyzing a collection of research papers, articles, and books. A researcher studying climate change could quickly analyze scientific papers to identify trends in sea level rise or assess the effectiveness of different mitigation strategies.
• Legal or regulatory compliance: Fine-tuning on legal documents can help automate contract review, flagging potential inconsistencies or areas of risk. This allows legal professionals to focus on higher-level tasks while ensuring compliance.
• Automated report generation: Automate the analysis of complex financial reports, extracting key performance indicators and generating summaries for stakeholders. This can save time and reduce the risk of errors compared to manual analysis.
• Content summarization and analysis: Summarize lengthy PDF documents, extract key insights, and analyze trends. For example, a market research team could analyze a collection of customer surveys to identify key themes and sentiment.
• Document comparison and version control: Compare different versions of a document to identify changes and track revisions. This can be particularly useful in collaborative environments where multiple authors contribute to a document.

Limitations

To learn more about document understanding requirements, seeDocument understanding.

Dataset format

ThefileUri for your dataset can be the URI for a file in a Cloud Storagebucket, or it can be a publicly available HTTP or HTTPS URL.

To see the generic format example, seeDataset example for Gemini.

The following is an example of a document dataset.

{"contents":[{"role":"user","parts":[{"fileData":{"mimeType":"application/pdf","fileUri":"gs://cloud-samples-data/generative-ai/pdf/2403.05530.pdf"}},{"text":"You are a very professional document summarization specialist. Please summarize the given document."}]},{"role":"model","parts":[{"text":"The report introduces Gemini 2.0 Flash, a multimodal AI model developed by Google DeepMind. The report positions Gemini 2.0 Flash as a significant advancement in multimodal AI, pushing the boundaries of long-context understanding and opening new avenues for future research and applications."}]}]}

What's next

• To learn more about the document understanding capability of Gemini models, see theDocument understanding overview.
• To start tuning, seeTune Gemini models by using supervised fine-tuning
• To learn how supervised fine-tuning can be used in a solution that builds agenerative AI knowledge base, seeJump Start Solution: Generative AIknowledge base.