[FT] Pipeline does not fully handle trust_remote_code to load dataset

[Sanahm]

Issue encountered

trust_remote_code is now mandatory to load dataset specifically with a script.

Solution/Feature

Add trust_remote_code as an input to create_custom_tasks_module() then in get_custom_tasks() and in Pipeline.pipeline_parameters (lighteval.tasks.registry).

class Pipeline:
    def _init_tasks_and_requests(self, tasks):
        with htrack_block("Tasks loading"):
            with local_ranks_zero_first() if self.launcher_type == ParallelismManager.NANOTRON else nullcontext():
                # If some tasks are provided as task groups, we load them separately
                custom_tasks = self.pipeline_parameters.custom_tasks_directory
                tasks_groups_dict = None
                if custom_tasks:
                    _, tasks_groups_dict = get_custom_tasks(custom_tasks, trust_remote_code=self.pipeline_parameters.trust_remote_code)
                if tasks_groups_dict and tasks in tasks_groups_dict:
                    tasks = tasks_groups_dict[tasks]

                ...

Maybe a more general fix could be to pass a generic **kwargs as an additional input to get_custom_tasks().

[clefourrier]

Hi! I'm quite sure we fixed this since, can you check if that's working for you from main?