What Counts as Diverse, Reliable Training Data?
APIGen's verification loop can tell whether a candidate works, but a verified dataset can still be narrow. This chapter examines how APIGen and xLAM define diversity through four kinds of function-calling request, then follows the less visible choices that determine what enters the library at all. The reader's task is to judge diversity along separate dimensions—call structure, API coverage, language variation, and reliability—and to see why a large synthetic dataset still depends on curated sources and capable generators.
Diversity is not one dimension
APIGen organizes its generated examples into four query styles. They are not four labels for the same difficulty. Each adds a different decision to the model's job.
A simple query provides one described API and expects one function call. The model must connect the request to the function and fill its arguments. If a user asks for the weather in one city and the available tool describes a weather lookup, the central decision is whether that tool and those arguments fit the request. This is the smallest useful unit of function calling, but it does not test selection among competing actions or coordination across requests.
A multiple query provides several APIs or functions and asks the model to choose the appropriate one. The difficulty is no longer just producing a valid call. The model must compare descriptions and reject plausible but irrelevant tools. A request about a film should not be routed to a weather function merely because both calls have a location-like parameter. Multiple examples test whether the model reads the action space rather than reaching for the first function that resembles a familiar pattern.
A parallel query contains several intentions that should be answered with multiple calls in one response. The same function may be invoked more than once with different arguments—for example, retrieving weather for Palo Alto and Paris together. The model must preserve all the requests, create the right number of calls, and keep the arguments associated with the correct target. This is a coordination problem even when every individual call is simple.
A parallel-multiple query combines both forms of choice. Several functions are available, and one or more of them may need to be called repeatedly. The model must select among tools and coordinate the resulting calls. A response can fail by choosing the wrong function, omitting one of several required calls, duplicating a request, or mixing arguments between calls. The category therefore tests a different capability from simply making a longer list of single calls.
These styles form a useful progression: choose one action, choose among actions, repeat an action for several intentions, then combine selection and repetition. They are not a universal scale of agent intelligence. A model can be strong on simple calls and weak on relevance detection, or accurate on multiple selection while failing to keep parallel arguments aligned. But the categories give the data and evaluation pipeline distinct questions to ask instead of treating every function call as the same task.
APIGen creates variation inside these categories through three samplers.