Adaptive stopping
AdaptiveStopping
dataclass
#
AdaptiveStopping(
max_responses: int | None = None,
min_confidence: float | None = None,
ambiguity_threshold: float | None = None,
)
Stops collecting responses on a compare datapoint once the outcome is clear.
Responses are collected until it is clear which option wins, until the two options
are so close that the datapoint is called ambiguous, or until max_responses is
reached. Clear pairs finish after a few responses; close pairs use more.
It is the default for compare job definitions. Pass it as responses_per_datapoint
to change a parameter, or assign it to RapidataJobDefinition.responses_per_datapoint
to switch an existing compare definition. Other job types raise ValueError.
Leave a parameter as None to use its default. A saved job definition always
carries the resolved values.
Parameters:
| Name | Type | Description | Default |
|---|---|---|---|
max_responses
|
int
|
The most responses collected per datapoint. Default: 100. Cost and progress count this many responses per datapoint. |
None
|
min_confidence
|
float
|
How certain the result must be before a winner is called, between 0 and 1 (exclusive). Default: 0.9. |
None
|
ambiguity_threshold
|
float
|
How close the two options must be before the datapoint is called ambiguous, as a normalized Gini impurity between 0 and 1 (exclusive). Higher is stricter: it needs a closer tie, so it costs more responses. Default: 0.64, which matches a 20%-80% split. |
None
|
Attributes:
| Name | Type | Description |
|---|---|---|
max_responses |
int | None
|
The response cap, or None for the default. |
min_confidence |
float | None
|
The confidence level, or None for the default. |
ambiguity_threshold |
float | None
|
The ambiguity threshold, or None for the default. |