Finally, how do you feel about the position taken in this note? It refers to the Imagenet dataset, which, like Laion, was distributed publicly only as a list of IDs (but was available as images to researchers on request).
@jeremyphoward
-
Legal implications of curated image datasets in AI training
By
–
And how about the "aesthetic" subsets? Are they likely to be treated differently because they're curated? (And they would have required accessing the image itself to do that curation.)
-
Google vs LAION: Legal Differences in Image Storage Liability
By
–
I'd love to better understand how Google, for instance, seems to be able to store and provide public links to images — but perhaps Laion can't? Are there some key differences that impact liability/legality?
-
Copyright implications of URL-based AI training datasets
By
–
very interesting- thanks for sharing! One thing not mentioned in the post that I'm curious about is that datasets like laion-5b don't actually contain copyrighted data directly, but are just lists of URLs. How does that impact potential claims, if at all?
-
Open Source Ethics: Responsibility Beyond Licensing
By
–
I'm referring to "Just because you are open source doesn't in my view support this lack of interest in doing the right thing".
-
Fine-tuning vs Full Training: Storage and Efficiency Trade-offs
By
–
Yeah we used to do that in earlier versions of the http://
fast.ai course, although it can eat up an awful lot of disk space pretty quick, so I don't bother any more! Fine-tuning the last few layers is generally fast enough anyway. -
Open Source Dataset Policies: Fair Treatment Concerns
By
–
Yeah you're probably right. (I hope they're making similar public statements about non open source use of that dataset too, and not just singling out the open source community.)
-
Hugging Face Dataset Removal Request Script Issues Clarified
By
–
So IIUC this thread has authors telling HF to remove a dataset which they don't actually host, but they provide a script that doesn't actually work, and they're asking HF to resolve that problem? I don't understand what problem needs to be addressed, or how.
-
HuggingFace Books3 Dataset Creation and Responsibility
By
–
It's a stretch to say HF created and maintained the books3 dataset. All they did was provide a script to allow others to download the dataset from a 3rd party server.
-
Julia Package Manager Similarities with Conda Compared
By
–
Actually Julia's pkg manager is nearly identical to conda. They both do this.