Can't download (some) models although they are in the hub

Beginners
1

Can’t download (some) models to pytorch, although they are in the hub (tried also the from_tf flag)

Error:
404 Client Error: Not Found for url: https://huggingface.co/umarayub/t5-small-finetuned-xsum/resolve/main/config.json

Models for example:
all of those models give 404 when trying to download them
[
“SvPolina/t5-small-finetuned-CANARD”, “Edwardlzy/t5-small-finetuned-xsum”, “Teepika/t5-small-finetuned-xsum”,
“HuggingLeg/t5-small-finetuned-xsum”, “V3RX2000/t5-small-finetuned-xsum”, “Teepika/t5-small-finetuned-xsum-glcoud”,
“VenkateshE/t5-small-finetuned-xsum”, “Wusgnob/t5-small-finetuned-xsum”, “HugoZhu/t5-small-finetuned-xsum”,
“Zazik/t5-small-finetuned-xsum”, “Paramveer/t5-small-finetuned-xsum”, “arkosark/t5-small-finetuned-xsum”,
“RamadasK7/t5-small-finetuned-squad”, “bochaowei/t5-small-finetuned-cnn-wei2”, “Kyaw/t5-small-finetuned-xsum”,
“ggosline/t5-small-herblables”,
]

Thise for example does work: “valhalla/t5-small-qa-qg-hl”,“mrm8488/t5-small-finetuned-quora-for-paraphrasing”

2

Looking at umarayub/t5-small-finetuned-xsum at main, there’s indeed no files in there.

There’s no config.json uploaded in that repo.

3

So basically it is corrupted and needs to be removed?
Isn’t there some check in place when uploading a model that it can be loaded, I have about twenty of those (and that is only in T5-small based models

4

Hi,

You can just add as many files as you’d like.

Assuming you are the owner of that repo on the hub, you can locally clone the repo (in a local terminal):

git lfs install
git clone https://huggingface.co/umarayub/t5-small-finetuned-xsum

You can then add files to that repo, and do git add ., git commit, and git push.

1 Like
5

Hey Niels thanks for the help.
Obviously, I am not the owner of those 20+ repositories I just want to load existing models in the hub. Are the models corrupted beyond repair? If so, how do we remove them from the hub and avoid corrupted models being uploaded in the future? If not, how can they be loaded (e.g. if only the config is missing, can’t we assume it is just like the default t5-small config)?

Also trying to supply a config doesn’t seem to change anything (but I might be doing it the wrong way.
Tried:

from transformers import AutoModelForSeq2SeqLM
AutoModelForSeq2SeqLM.from_pretrained(‘SvPolina/t5-small-finetuned-CANARD’, config=‘t5-small’)

For future reference, models that just fail loading (from pytorch, with and without from_tf) I assume jax ones could be loaded in some way ASP, the others seem to be problematic.

jax_models = [“aqj213/t5-small-pisa-state-only-finetuned”, “shivam12/t5_small_pubmed”]
not_working = [
“SvPolina/t5-small-finetuned-CANARD”, “Edwardlzy/t5-small-finetuned-xsum”, “Teepika/t5-small-finetuned-xsum”,
“HuggingLeg/t5-small-finetuned-xsum”, “V3RX2000/t5-small-finetuned-xsum”, “Teepika/t5-small-finetuned-xsum-glcoud”,
“VenkateshE/t5-small-finetuned-xsum”, “Wusgnob/t5-small-finetuned-xsum”, “HugoZhu/t5-small-finetuned-xsum”,
“Zazik/t5-small-finetuned-xsum”, “Paramveer/t5-small-finetuned-xsum”, “arkosark/t5-small-finetuned-xsum”,
“RamadasK7/t5-small-finetuned-squad”, “bochaowei/t5-small-finetuned-cnn-wei2”, “Kyaw/t5-small-finetuned-xsum”,
“ggosline/t5-small-herblables”,“Luckyseeker/t5-small-finetuned-xsum”,“umarayub/t5-small-finetuned-xsum”,
“yougang/t5-small-finetuned-xsum”,“xikoto/t5-small-finetuned-xsum”, “vhvk99/t5-small-finetuned-xsum”,
“tsosea/t5-small-finetuned-xsum”,“tharik/t5-small-finetuned-xsum”,“malay-huggingface/t5-small-abstractive-summarization-bahasa-cased”, “Alifarsi/t5-small-finetuned-xsum”,
“heejun/t5-small-finetuned-xsum”,
“MHJ/t5-small-finetuned-xsum”,
“kroshan/t5-small-finetuned-xsum”,
“E312/t5-small-finetuned-xsum”,
“knkarthick/t5-small-finetuned-xsum”,
]

6

the hf hub hosts user- and organization-generated content so we don’t guarantee all models work, but depending on your use case you could use tag filters to filter for models that are more likely to be what you’re looking for:

https://huggingface.co/api/models?filter=t5&sort=downloads&direction=-1

will filter for models detected to be of type T5, and sort them by downloads. (downloads is one possible indication of whether a model is “good”)

1 Like
7

I really don’t understand, you mean you think it is a desirable trait that there are model names, but their content is completely empty? without any file in them?

I assume the one uploaded them didn’t notice they did, or failed to upload, or got stuck in the middle. Why would it be good for users to have hosted “models” with only the .gitattributes file in them? (which is the case of those models, e.g. knkarthick/t5-small-finetuned-xsum at main )
Promise that if you understand those models are totally empty and think it is the expected behavior I will stop asking. :slight_smile: