Training README datasets table: text column and id column

[guynich]

Training README datasets table: text column and id column

[guynich]

For VoxPopuli I find raw_text to contain empty strings and think that is why my pseudo-labelling script failed after hours of compute with a ValueError("one or more references are empty strings").

For facebook/voxpopuli train split I find 5,463 empty "raw_text" strings of the 182,482 examples. Each empty "raw_text" string has a corresponding non-empty "normalized_text" string.

[guynich]

Copied from #98
When pseudo-labelling the Voxpopuli dataset the "raw_text" (needed for option --text_column_name) may be an empty string for some examples - see HF dataset model card here for an empty "raw_text" example.

Question: how do I check which text name ("raw_text" or "normalized_text") was used when creating the pseudo-labelled datasets on HF, such as https://huggingface.co/datasets/distil-whisper/voxpopuli ?

[sanchit-gandhi]

Hey @guynich - the provided transcriptions in the original VoxPopuli dataset are only used for computing the WER in the pseudo-labelling and distillation scripts. Since the WER is computed on normalised transcriptions, you can safely use the "normalized_text" column in the dataset, which is what was done for the Distil-Whisper datasets.

If you do decide to use the un-normalised (raw) text column, you should filter out any empty transcriptions from your dataset using a raw_datasets.filter method, e.g. as done here:

distil-whisper/training/run_distillation.py

Lines 1224 to 1236 in b948d02

	# 10.6: Filter training data with labels longer than `max_label_length`
	def is_labels_in_length_range(labels):
	return 0 < len(labels) <= max_label_length
	filter_by_labels_fn = partial(
	vectorized_datasets.filter, function=is_labels_in_length_range, input_columns=["labels"]
	)
	if accelerator.is_main_process:
	vectorized_datasets = (
	filter_by_labels_fn(num_proc=num_workers, desc="filtering train dataset")
	if not data_args.streaming
	else filter_by_labels_fn()
	)

[guynich]

Thank you for the helpful comment and for the fix #102. Closing.