search

google-gemini / generative-ai-python

The official Python library for the Google Gemini API
https://pypi.org/project/google-generativeai/
Apache License 2.0
1.64k stars 326 forks source link

Safety filter is too strong today #65

Closed xiaohk closed 1 year ago

xiaohk commented 1 year ago

Description of the bug:

Hello, I am using the PaLM API through Python / JS and MakerSuite. I noticed that the safety filter has become unexpectedly strong recently. My very benign inference requests were all responded with {reason: "OTHER"}. Is there a way to change my safety setting to get the result?

Here is an example in MakerSuite (I have already lowered the safety filter thresholds, see screenshot 2):

image image

Actual vs expected behavior:

I expect to get result "comment vas-tu ?" instead of an empty response with OTHER reason. I believe PaLM 2 API used to give me this correct response two days ago.

Any other information you'd like to share?

No response

xiaohk commented 1 year ago

I used this MakerSuite prompt.

xiaohk commented 1 year ago

I believe PaLM 2 is designed to support better translation, so it would be great if PaLM 2 API users can use the model for translation tasks. See section 4.5 in the PaLM 2 technical report.

An explicit design choice of PaLM 2 is an improved translation capability.

image
williamito commented 1 year ago

Hi there; this PaLM API only supports English at the moment (https://developers.generativeai.google/models/language#model_attributes) so translation use cases are not supported.

xiaohk commented 1 year ago

I see, thank you for the reply. @williamito