Wrong exception handling when loading dataset from local disk

[ganler]

alignment-handbook/src/alignment/data.py

Lines 216 to 221 in 606d2e9

	try:
	# Try first if dataset on a Hub repo
	dataset = load_dataset(ds, ds_config, split=split)
	except DatasetGenerationError:
	# If not, check local dataset
	dataset = load_from_disk(os.path.join(ds, split))

Actual exception is ValueError:

[rank5]: Traceback (most recent call last):
[rank5]:   File "run_sft.py", line 251, in <module>
[rank5]:     main()
[rank5]:   File "run_sft.py", line 86, in main
[rank5]:     raw_datasets = get_datasets(
[rank5]:   File "miniconda3/envs/handbook/lib/python3.10/site-packages/alignment/data.py", line 169, in get_datasets
[rank5]:     raw_datasets = mix_datasets(
[rank5]:   File "miniconda3/envs/handbook/lib/python3.10/site-packages/alignment/data.py", line 218, in mix_datasets
[rank5]:     dataset = load_dataset(ds, ds_config, split=split)
[rank5]:   File "miniconda3/envs/handbook/lib/python3.10/site-packages/datasets/load.py", line 2570, in load_dataset
[rank5]:     raise ValueError(
[rank5]: ValueError: You are trying to load a dataset that was saved using `save_to_disk`. Please use `load_from_disk` instead.

Dataset version:

❯ pip show datasets
Name: datasets
Version: 2.19.1
Summary: HuggingFace community-driven open-source library of datasets
Home-page: https://github.com/huggingface/datasets
Author: HuggingFace Inc.
Author-email: [email protected]
License: Apache 2.0
Location: /home/ec2-user/miniconda3/envs/handbook/lib/python3.10/site-packages
Requires: aiohttp, dill, filelock, fsspec, huggingface-hub, multiprocess, numpy, packaging, pandas, pyarrow, pyarrow-hotfix, pyyaml, requests, tqdm, xxhash
Required-by: alignment-handbook, evaluate, trl

Also tried the latest 2.19.2 and got the same error. Need to broaden the exceptions to capture.

[alvarobartt]

Hi here @ganler, thanks for reporting! Do you want to open a PR to fix the data loading handling? Otherwise, feel free to ping us and we can have a look at it, but as you're pointing out, the following should do the work:

-     except DatasetGenerationError:
+     except ValueError:

Thanks in advance! 🤗

[ganler]

Thanks! That's how I fixed it temporarily. Feel free to fix it on your side.

[xiyang-aads-lilly]

@alvarobartt seems that this issue has not been fixed in the repo

alignment-handbook/src/alignment/data.py

Line 219 in 01f29c1

except DatasetGenerationError:

.

more interesting is that if we use "train" or "test" as splits, it can load the data that saved use save_to_disk function in wrong way. So change to ValueError is just a temp solution. Any suggestion on a better way to handle this problem?
e.g.,

raw_datasets.save_to_disk("data")
datasets.load_dataset("data", split="train")
# output:
Dataset({
    features: ['_data_files', '_fingerprint', '_format_columns', '_format_kwargs', '_format_type', '_output_all_columns', '_split'],
    num_rows: 1
})