content-learning-examples

To generate SFT training data from PDFs, use the document-grounded path to extract Q&A pairs directly from source text chunks. Answers are strictly grounded in the provided document content.

Creating a question answering dataset without documents uses a topic-driven approach that decomposes domains into subtopics and applies web search to label answers for SFT dataset generation.

Yes, after generating Q&A pairs, the pipeline supports dataset linting, quality filtering, and train/test splitting to format training examples for hosted supervised fine-tuning runs.

Yes, it supports production dataset creation for domain teaching and knowledge injection, including generating survival-field instructions and clinical nutrition Q&A extracted from medical textbooks.

Topic decomposition breaks a domain seed into targeted subtopics, then uses a web search labeler to find and validate answers for SFT Q&A pairs when no documents are provided.

Use document ingestion when you have source PDFs for grounded answers, and use web search when you only have a topic seed and need web-validated answers for your Q&A training examples.