Training / Standard term
Pretraining
Training a model on a broad objective and large dataset before adapting it to narrower instructions, preferences, or tasks.
Language-model pretraining commonly predicts tokens from text and code; multimodal systems may also train on image, audio, or video objectives. Dataset size, architecture, compute, filtering, and objective vary widely. The resulting weights support later behavior, but capabilities cannot be assigned cleanly to one stage because post-training, tools, and inference procedures can add, expose, suppress, or redirect performance.
Builder example
Providers often disclose only broad data categories, so builders usually cannot verify domain coverage by reading a training manifest. Evaluate the deployed model and tool configuration on representative domain tasks, current knowledge, failure cases, and source-use requirements.
Common confusion: Pretraining is not merely reading facts into a database, and a base model is not limited to literal autocomplete interfaces. Post-training changes behavior, but the boundary between capabilities learned at each stage is empirical.