Many models in the < 1.5B weight class are actually > 70B

[surya-narayanan]

Hi, sorry if my understanding is unfounded, but it looks like most of the models that are displayed to be < 1.5B, are actually modifications of 70B models.

I'm explicitly looking for good, small, LLMs, so this feature, if it worked, would be extremely useful for me.

Please let me know if I'm misunderstanding something.

[osanseviero]

cc @clefourrier

[clefourrier]

Hi! Thank you for your message :)

This is linked to two issues:

• when we can't get a number of parameters from the hub or the model name, we select 0 as the number of parameters. ~I'll try to add another button to remove these models asap.~ Edit: fixed
• we also have an issue with the way weights are computed in safetensors for quantized models that you can follow here.

Best way to solve this (while the team is working on it) is to display the precision field and parameter number, and only look at non quantized models with a number of params > 0.