Text Generation
Transformers
Safetensors
falcon
text-generation-inference

I think the Falcon-180B web app is having a stroke

#26
by andreped - opened

I just tried the open web app that was made openly available at https://huggingface.co/blog/falcon-180b#demo

Sadly, it seems to just be outputting nonsense. Here is a very simple example of what it produces based on the example prompts:

Skjermbilde 2023-11-16 114524.png

I was trying to use it to draw some information from a rather large JSON document, which worked perfectly fine with the other, smaller Falcon LLMs, but for this web app, the output seems downright bad. It does not even make sense. Maybe there is something that broke?

For an industry project, it would be of great interest to test this model, but right now I have no easy way to test it.

Have you tried this web app lately?

cc: @philschmid @osanseviero

Just observed that the web app is hosted in a separate space here:
https://huggingface.co/spaces/tiiuae/falcon-180b-demo/

There also seems to be multiple people reporting the same (see here).

So perhaps we could move the discussion there instead.

Might be an idea to keep this Discussion open, until the problem has been resolved, to avoid others making the same thread here.

Sign up or log in to comment