diff --git a/chatsky/utils/db_benchmark/basic_config.py b/chatsky/utils/db_benchmark/basic_config.py index 68d9c1006..edd0cc620 100644 --- a/chatsky/utils/db_benchmark/basic_config.py +++ b/chatsky/utils/db_benchmark/basic_config.py @@ -96,9 +96,9 @@ class BasicBenchmarkConfig(BenchmarkConfig, frozen=True): Number of times the contexts will be benchmarked. Increasing this number decreases standard error of the mean for benchmarked data. """ - from_dialog_len: int = 300 + from_dialog_len: int = 25 """Starting dialog len of a context.""" - to_dialog_len: int = 311 + to_dialog_len: int = 50 """ Final dialog len of a context. :py:meth:`~.BasicBenchmarkConfig.context_updater` will return contexts @@ -177,27 +177,25 @@ def context_updater(self, context: Context) -> Optional[Context]: basic_configurations = { "large-misc": BasicBenchmarkConfig( from_dialog_len=1, - to_dialog_len=50, + to_dialog_len=26, message_dimensions=(3, 5, 6, 5, 3), misc_dimensions=(2, 4, 3, 8, 100), ), "short-messages": BasicBenchmarkConfig( - from_dialog_len=500, - to_dialog_len=550, message_dimensions=(2, 30), misc_dimensions=(0, 0), ), "default": BasicBenchmarkConfig(), - "large-misc--long-dialog": BasicBenchmarkConfig( - from_dialog_len=500, - to_dialog_len=550, + "large-misc-long-dialog": BasicBenchmarkConfig( + from_dialog_len=50, + to_dialog_len=75, message_dimensions=(3, 5, 6, 5, 3), misc_dimensions=(2, 4, 3, 8, 100), ), "very-long-dialog-len": BasicBenchmarkConfig( context_num=10, - from_dialog_len=10000, - to_dialog_len=10050, + from_dialog_len=1000, + to_dialog_len=1050, ), "very-long-message-len": BasicBenchmarkConfig( context_num=10,