fix(map): fix progress bar exceeding total when load_from_cache_file=False

[Nitin-Rajasekar]

Summary

When load_from_cache_file=False, the progress bar in Dataset.map() displayed a count that exceeded the dataset size (e.g. "800 examples" for a 400-row dataset).

The bug: pbar_initial was computed using len(existing_cache_files), which counts cache files on disk regardless of whether they would be used. With load_from_cache_file=False, those files exist but are ignored, yet the bar was initialized to the full dataset size as if all work was already done. The actual processing then added on top of that, pushing the counter past the total and causing tqdm to drop the percentage display entirely.

Fix: use num_shards - len(unprocessed_kwargs_per_job) instead, which is 0 when load_from_cache_file=False, so the bar starts and counts correctly.

Issue

Fixes #8167

Local verification

py -3.14 -m pytest tests/test_arrow_dataset.py -k "test_map_load_from_cache_file_false_progress_bar_starts_at_zero" -v

2 passed

Risk

Low risk, a one-line change to how pbar_initial is computed. No behaviour changes outside of the progress bar display.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.