record_checksums default changed True -> False in 4.8.5, not in the release notes
What changed
Between 4.8.4 and 4.8.5, checksum recording flipped from on to off by default in two places. From the sdists:
4.8.4 src/datasets/download/download_manager.py:80 record_checksums=True
4.8.5 src/datasets/download/download_manager.py:80 record_checksums=False
4.8.4 src/datasets/utils/info_utils.py:80 record_checksum: bool = True
4.8.5 src/datasets/utils/info_utils.py:80 record_checksum: bool = FalseWhy I am asking rather than reporting
The 4.8.5 release notes do not contain the words checksum, integrity, or verification. The closest entry is:
Don't include files list DatasetInfo (and remove old stuff) by @lhoestq in #8128
whose stated motivation is that DatasetInfo was making the viewer's Mongo documents too large, with checksum recording described as cleanup alongside it. So this reads to me like a deliberate change that came in as a side effect of an unrelated fix, rather than a decision anyone announced.
I may well be misreading it, which is why this is a question.
The reason it matters to someone downstream
DatasetInfo.download_checksums stops being populated, so anything relying on it for provenance or reproducibility silently gets nothing rather than an error. It went out in a patch release, where the version number promises no behaviour change at all.
What would help
Either a line in the release notes or a note in the docs confirming that checksum recording is now opt-in, and ideally a sentence on what the recommended replacement is for verifying downloaded files. If it was not intended to change the default, that is worth knowing too.
How I found it
A tool I maintain diffs default values of public callables between two releases and grades by whether the version number promised anything. This came out of a sweep across 1,253 consecutive patch releases; it was one of a small number where the changed default looked consequential.
Source: huggingface/datasets