Apologies if this has been discussed before, I didn’t see anything about it searching the forum.
Problem: The line graph at the top of (every?) model card showing “Downloads last month” seems for some reason to always have a zero or near-zero value for the last data point.
In other words, they all take a “nose dive” or look like a stock market crash; for every model, it depicts the idea that “this model is no longer popular among anyone/everyone stopped using it, regardless of whatever trend may may be apparent for the other days.
I personally didn’t realize until just now that this is merely an effect of the graph, and not an indicator of the trend. I’ve passed up using many models seeing this, thinking “looks like it’s not downloaded anymore, there must have been a better model that people all know about and are downloading instead!”.
My guess is that it’s merely just graphing downloads for discrete time periods, but that the algorithm adds an incomplete (and apparently continuously extending [i.e., not every 24hrs, perhaps hourly or instantaneously?], consequently always making the graph dive at the end.
I would like to suggest, that at the very least, the graph algorithm be updated so this doesn’t happen.
At a very minimum, I think this could perhaps be as simple as only showing data up until 24 hours ago (e.g. do not show the last 24 hour’s downloads, only the 30 days before it). From there, I can imagine a myriad of ways that could eliminate the “last-minute dive” and simultaneously perhaps provide more useful or in-depth information about the model’s popularity… ideally, it would be awesome if his was just an interactive widget: Perhaps it loads user’s preference, but can be made larger (“theater mode”), and then adjustable like setting the time period and the metrics, and compare to others. For example, effectively choose to “view downloads (ideally by file?) together with online inference usage for the past 6 months, and overlay the same timeline and metrics for some other model.”
Or, maybe provide the raw data (or do you already, sorry I checked but admittedly didn’t spend too much time doing so) to allow full customized evaluation. My ultimate goal would be to basically look at trending models, but where “trending” is specifically tailored, vetted, and pruned to provide results more relevant to me personally.
Apologies, this is much longer than I meant it to be.
Just wanted to share my 2 cents on this particular item: especially the graph dive thing… I greatly appreciate everything you all do, regardless.
Cheers! 