How LLM training data is sourced, labeled and quality-checked at each training stage

Tell me what you know about gathering and labeling the data that large language models are trained on. ```hint Split by stage A model learns from different data at different stages of training, in very different volumes. Ask which stages need a person to label anything at all, and what that person is actually judging in each. ``` ### Clarifying Questions - Should I cover every stage (pretraining, supervised fine-tuning, preference tuning, evaluation sets) or focus on the stage where human annotation matters most? - Is the context a general-purpose foundation model, or an enterprise team adapting an existing model to its own domain? - Do you want this as a product opportunity, such as tooling for teams that need curated and labeled data, or as how a model team should run its own data pipeline? ### What a Strong Answer Covers - How data needs differ across pretraining, supervised fine-tuning, preference tuning and evaluation, both in volume and in kind - Sources of data (web crawls, licensed corpora, human-written examples, model-generated synthetic data, user feedback) and the legal, privacy and consent limits on each - How annotation is run: guidelines, annotator expertise, quality control and agreement measures, and cost - Curation steps that matter as much as volume: deduplication, filtering, keeping evaluation data out of training, and choosing the data mix - A product view: where teams building or adapting models struggle with data, and what would help them ### Follow-up Questions - When is synthetic data from a stronger model a good substitute for human annotation, and when does it go wrong? - How would you measure whether a new batch of annotations actually improved the model? - A customer wants to fine-tune a model on its support tickets. What data and labeling plan would you recommend?