That's really cool ❤️
Out of curiosity, what use case did you build this dataset for?
Any information we could expose natively from the API, that would make your lives easier?
Thanks for sharing! 🤗
Thanks for liking Julien. The inspiration behind this dataset for us is to solve the problem of exploration & helping in model selection.
Using the info on type of inputs (image or text), we are able to filter where models = image->image and search for "Sketch" models for example.
Then we're able to run all the sketch models and cache the outputs, so that horizontal comparison between similar spaces becomes easy, and we can select the ones which we like and explore them further. Caching also saves compute time and cost, since we don't need to run multiple heavy models again. Using this dataset, we're also thinking of creating spaces similar to RL leaderboard space where folks can see cached output across "Sketch", "Toonify", "Zombify", "TextToImage" etc categories using a pre-selected set of images / audio / text prompts, and hopefully it helps others in choosing the right model (subjectively) without having to run everything.
And as an added benefit, we get rid of simply static spaces which aren't of use to us eg: https://huggingface.co/spaces?sort=modified&search=animation most of spaces here are built for personal use and we can mark them as "not interested" in our local copy of the dataset.
Regarding natively exposing info from API, I have a suggestion. We couldn't find any info on "Past" Spaces of the week. Maybe a dedicated HallOfFame page, or exposing this badge on every space can also serve as a soft indicator of good spaces? This would be similar to "Expert approved" apps in app stores.
super interesting! cc @osanseviero
We've thought about HallOfFame as well! We'll think more about this and let you know!