updated val batch size for the environments

    data.train_files=data/navigation-vision-debug/train.parquet \

    data.val_files=data/navigation-vision-debug/test.parquet \

    data.train_batch_size=2 \

    data.val_batch_size=2 \

    data.max_prompt_length=1024 \

    data.max_response_length=128 \

    data.max_trajectory_length=1800 \

    data.train_files=data/sokoban-text-debug/train.parquet \

    data.val_files=data/sokoban-text-debug/test.parquet \

    data.train_batch_size=16 \

    data.val_batch_size=128 \

    data.max_prompt_length=1024 \

    data.max_response_length=128 \

    data.max_trajectory_length=1800 \

    data.train_files=data/sokoban-vision-debug/train.parquet \

    data.val_files=data/sokoban-vision-debug/test.parquet \

    data.train_batch_size=16 \

    data.val_batch_size=128 \

    data.max_prompt_length=1024 \

    data.max_response_length=128 \

    data.max_trajectory_length=1800 \

  max_response_length: 512

  max_trajectory_length: 3072 # vagen only

  train_batch_size: 1024

  val_batch_size: null # DEPRECATED: Validation datasets are sent to inference engines as a whole batch, which will schedule the memory themselves

  val_batch_size: 1024 

  return_raw_input_ids: False  # This should be set to true when the tokenizer between policy and rm differs

  return_raw_chat: False

  shuffle: True

            dataset=self.val_dataset,

            # Validation datasets are sent to inference engines as a whole batch,

            # which will schedule the memory themselves.

            batch_size=len(self.val_dataset),

            batch_size=self.config.data.val_batch_size,

            num_workers=8,

            shuffle=False,

            drop_last=False,

            collate_fn=collate_fn)

        assert len(self.train_dataloader) >= 1

        assert len(

            self.val_dataloader

        ) == 1, "Validation dataloader must have a single batch, which inference engines will schedule the memory themselves."

        # assert len(

        #     self.val_dataloader

        # ) == 1, "Validation dataloader must have a single batch, which inference engines will schedule the memory themselves." # for agent training we still use val batch size

        print(f'Size of train dataloader: {len(self.train_dataloader)}')