Bug: AttributeError when using JSON/CSV/Parquet file datasets

[Basster]

Issue Description

When running guidellm run with a json_file, csv_file, or parquet_file data source, the benchmark fails with:

Traceback (most recent call last):
  File ".../guidellm/data/preprocessors/mappers.py", line 192, in datasets_mappings
    if dataset.info and dataset.info.dataset_name
       ^^^^^^^^^^^^
AttributeError: 'DatasetDict' object has no attribute 'info'

Root Cause

The file-based dataset deserializers (JSONFileDatasetDeserializer, CSVFileDatasetDeserializer, ParquetFileDatasetDeserializer) in .venv/lib/python3.12/site-packages/guidellm/data/deserializers/file.py are calling load_dataset() directly, which returns a DatasetDict object for these file types.

However:

1. The type annotations specify -> Dataset as the return type
2. Downstream code in guidellm/data/preprocessors/mappers.py:192 expects a Dataset with an .info attribute
3. DatasetDict is dict-like and doesn't have an .info attribute - individual splits like DatasetDict['train'] do

Expected Behavior

The benchmark should complete successfully, processing the dataset file correctly.

Actual Behavior

The benchmark fails during dataset preprocessing initialization with AttributeError: 'DatasetDict' object has no attribute 'info'.

Fix

Modified the three deserializers to extract the 'train' split from the DatasetDict when it's returned:

1. Added DatasetDict to imports: from datasets import Dataset, DatasetDict, load_dataset
2. Changed each deserializer to check if the result is a DatasetDict and extract the train split:

   dataset = load_dataset("json", data_files=str(path), **config.load_kwargs)
   if isinstance(dataset, DatasetDict):
       dataset = dataset["train"]
   return dataset

This ensures the deserializers return a Dataset object as typed, matching the behavior of other deserializers like TextFileDatasetDeserializer.

I've prepared a patch file which got it working on my machine: guidellm_fix.patch

Maybe I find the time to open a PR, but currently I don't thinks so. Sorry 😬

[Pragadeesh122]

Thanks for the detailed root-cause analysis @Basster — picking this up. I've extended the fix to also cover the arrow_file and tar_file deserializers, which hit the same DatasetDict issue via the same load_dataset path. PR incoming.