Hugging Face has released Datasets 5.2.0, adding faster streaming decode for image, audio and video folders alongside fixes for data handling, exports and distributed use. One notable security-related change removes loading of pickled pandas dataframes, according to the release notes.
Hugging Face Watch analysis
What happened
The Datasets 5.2.0 release notes list changes across dataset features, loaders and storage. Media-folder streaming decode is now faster, and the mesh feature and meshfolder builder accept .gltf files.
The release also fixes several concrete snags: pushtohub dropping rows when the number of processes exceeds the number of shards; batch accumulation triggering full downloads; saving batches containing more than 2GB of image or audio files; and distributed dataset splitting that could give every rank all the data. The notes also say pickled pandas dataframe loading has been removed, citing CVE-2026-57845.
Why it matters
For people building or preparing AI datasets, faster streaming across common media types and support for .gltf meshes make more varied data easier to work with. The bug fixes address failures that could waste time or produce the wrong data distribution, rather than merely tidying the changelog.
The pickle-loading change is worth noticing, but the release notes alone do not explain the vulnerability’s impact or affected versions. Treat it as a security-relevant change, not a full security advisory.
Our read
This is a useful maintenance release with several tangible wins for dataset builders, especially those handling large media batches or distributed workloads. If you use Hugging Face Datasets, check the notes for fixes that match your pipeline and review the pickle-loading change before upgrading anything security-sensitive. A version number is not a migration plan, even when the changelog is doing solid work.
What to watch
- Whether the faster media-folder streaming improves throughput for your datasets and setup.
- Whether the pickle-loading removal affects any existing workflows, and what further security details are published.
- Whether the fixes resolve issues in large-file saving, sharded uploads or distributed splits.
Discussion spark: For dataset tooling, which deserves priority: smoother support for more media types, or fixes that prevent pipelines from silently handling data incorrectly?
Sources and evidence
- 5.2.0 (9 November 2026, 21:40 UTC)
not affiliated with or endorsed by Hugging Face