Privacy of transformers models?

[burdonmark]

Hi,

It's not clear to me if queries fed into transformers via huggingfaceR are ultimately reported back to HuggingFace. It seems there is a telemetry setting in the python package that can be turned off:

https://github.com/huggingface/transformers/blob/d211a84aca8a8513a599171173c29bc8d3061fba/src/transformers/utils/hub.py#L106

DISABLE_TELEMETRY = os.getenv("DISABLE_TELEMETRY", False) in ENV_VARS_TRUE_VALUES

Definitely possible I am misunderstanding this but there doesn't seem to be a clear statement online about whether queries to the open-source models are reported.

[jpcompartir]

Hi, apologies - package is not currently maintained.

Is the issue here around using the Hugging Face hub i.e. do Hugging Face know which models you search for, or downloading a model such as an LLM and then using it for inference?

As far as I know, once you download the model and run it on a local GPU you don't need to worry about Hugging Face taking your data. Take that with a pinch of salt, though.

From the doc you linked, which is a part of the HF hub, there's this section:

` Sends telemetry that helps tracking the examples use.

Args:
    example_name (`str`): The name of the example.
    *example_args (dataclasses or `argparse.ArgumentParser`): The arguments to the script. This function will only
        try to extract the model and dataset name from those. Nothing else is tracked.
    framework (`str`, *optional*, defaults to `"pytorch"`): The framework for the example.`

I don't know of anything similar for actually running inference locally/privately