Tutorial for Shapley Value Sampling likely needs update

[ayan-cpath]

📚 Documentation

In the tutorials for perturbation based attribution for HuggingFace CasualLMs, in the block 12 that demonstrates how to use SHAP values, the parameter skip_tokens is being passed:

Running that code gives the error - TypeError: ShapleyValues.attribute() got an unexpected keyword argument 'skip_tokens':

Neither the API doc nor the source has the attribute method has parameter skip_tokens, likely removed in one of the more current versions of the package.

Thanks for the great package and documentation.

[sarahtranfb]

What version of Captum are you on? The tutorials are for the latest version; based on the screenshot provided, you're on a previous version where LLMAttribution.attribute() doesn't have skip_tokens as a parameter, and therefore it's being passed to ShapleyValues.attribute() as part of **kwarg, though ShapleyValues never supported skip_tokens.

Based on the code snippet, please remove skip_tokens from LLMAttribution.attribute() or update to a newer version.

[ayan-cpath]

What version of Captum are you on? The tutorials are for the latest version; based on the screenshot provided, you're on a previous version where ...

Could be the case. I installed captum using pip install captum and it installed 0.7.0, which looks like the latest released version on here on github.

[ayan-cpath]

I can likely do something along the lines of pip install --upgrade git+https://github.com/pytorch/captum.git and get the latest one, but then, it would be great to get a new stable release or update the doc to mention this.

[sarahtranfb]

Ah yeah it's been a little over a year since the last package release, though we're actively trying to get one out soon. I'll leave this open and come back in a month if we still don't have a release to update the website and notebook to call out that a manual installation is required.