A training dataset is a collection of examples used to fit a model’s parameters and learned patterns. For large language models, examples can include tool definitions paired with the calls and arguments that correctly answer a request. section 3.2.2
Tool-use training examples teach a model to:
- Recognize when a request requires a tool. section 3.2.2
- Select an appropriate tool from the available options. section 3.2.2
- Extract parameter values and format a structured function call. section 3.2.2
Tool-call example
For “Please set the living room blinds to 75% openness,” a training example pairs the request with set_blind_openness(room='living room', openness_percentage=75). Repeated patterns support recognition, selection, argument extraction, and formatting. section 3.2.2
Training data can also supply factual knowledge: web content such as Wikipedia and reference materials may let a model answer documented questions without external tools. That knowledge remains unreliable for current or precise information because the model cannot reliably distinguish accurate recall from confabulation. section 2.4.3
