Is there a way to extract a model's download stats (e.g., last 30 days) in times series format?

[frr717]

Is your feature request related to a problem? Please describe.
Currently, I can only find code like below to get a static data point (last 30 day download count from today):
info = model_info("bert-base-uncased") model_info(info.modelId).downloads

Describe the solution you'd like
I wonder whether huggingface can provide methods with an input specifying the date? such as

model_info(info.modelId).get_downloads('20240131')

Describe alternatives you've considered
currently no...
I appreciate any help from all of you!

[Wauplin]

Hi @frr717, thanks for your interest. There is currently no way to get this data as time series. The only information you can get is the downloads in the last 30 days and overall downloads. What would be your use case for a timeseries format?

[frr717]

Hi @frr717, thanks for your interest. There is currently no way to get this data as time series. The only information you can get is the downloads in the last 30 days and overall downloads. What would be your use case for a timeseries format?

Thank you for you reply.

I am in a research project that needs to use this times series to conduct some regression analysis on companies that those models belong to. Hence I am interested to know whether your team has a plan to implement it?
Thanks!

[frr717]

Hi @frr717, thanks for your interest. There is currently no way to get this data as time series. The only information you can get is the downloads in the last 30 days and overall downloads. What would be your use case for a timeseries format?

BTW, I want to kindly ask you another questions regarding the image (a SVG element in the html of a model, such as [https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5] on the right top corner, besides the "Downloads last month":

What is the frequency of the data points in the image?
Take this model on the above link as an example: it was created on 2024-05-19. So, does this mean that the line in the SVG represents each DAY's last-30-day downloads since its creation time?
Thank you!

[julien-c]

it's each day in the last 30 days

[frr717]

it's each day in the last 30 days

thank you!

[frr717]

it's each day in the last 30 days这是过去 30 天内的每一天

hi, @julien-c
The data points on this image has been compressed to the range [0,100].

Could you kindly tell me the formula it uses?

Thank you!

[julien-c]

0 means 0 download, ie. we don't move the origin.

So yes, you can get the daily downloads from the last30days total + the graph. It's a bit hacky but it'll work.

[pilipentseva]

Hi @julien-c @Wauplin , I am research assistant at LSE for professor Jonathon Hazell. We actually encountered the same research need in Economics project as the one above. In our set up, it would be great to have some proxy for AI models frequency of use in regressions for macroeconomics research at daily frequency historically. And it looks like it the best (if only) source available. It is probably a very brave and long shot and I understand you dont have this feature publicly, but may it be somehow possible to have some discussion whether this data can be available for some models for research needs (we can make data confidential etc and only show regressions, cite Hugging Face in the research, share research design).

[Wauplin]

Hi @pilipentseva, thanks for reaching out. Unfortunately we do not provide this level of detail. The best dataset we have for research purposes is this one https://huggingface.co/datasets/cfahlgren1/hub-stats from @cfahlgren1. It does not provide as much granularity as you would like but it is updated regularly and always available. Go to the Data Studio tab to inspect the data. Hope this can help you move forward in your research 🤗