Default model selection about huggingface.js HOT 3 CLOSED

If it's possible that the default model changes over time and hf.js looks it up dynamically in /api/tasks, user's code that doesn't specify the model may suddenly break because of a different model response.
Also, in some tasks the default model doesn't make sense, e.g. tabular regression & classification.

from huggingface.js.

@vvmnnnkv , you are right about the fact that some tasks a default model doesn't make sense for a real use case. It's should be more seen as a way to discover the API itself. This feature is still expected to be useful for popular (and generic) tasks that like text-generation/image-generation, feature extraction, image classification/segmentation,...

About reproducibility, what we do in huggingface_hub (the Python client) is to log a message to recommend to the user to explicitly set model=*** as the recommended model might change without prior notice. The goal of having dynamic recommended models is to be able to suggest better models as they come without having to make releases in each client (JS and Python).

In the end, recommended models are mainly made for quick prototyping while we encourage to set models explicitly for more production-ready use cases. Also note that the free Inference API hosted on the Hugging Face Hub is already not really meant to be production-ready or expect high usage.

from huggingface.js.

cc @Wauplin ⬆️

from huggingface.js.