Hugging Face's logo

google
/
gemma-4-12B-it

Any-to-Any
Transformers
Safetensors
gemma4_unified
image-text-to-text
Eval Results
Model card Files
xet
 Community
40

Gemma 4 12b bias affraid me

#39
by Alekto - opened
Discussion
Alekto

Has anyone encountered this model's bias?

As soon as it receives some information (including self-generated and false), it will not deviate from this path.

Simple test:

  1. if you ask "Who won USA 2024 election?", model asked D. Trump and any other about election process.
  2. but if you ask "What is your actual knowledge cutoff?" and the model answers "January 2024" It will refuse to answer about everything after, even if you give arguments why this is a mistake

I noticed this phenomenon on another question, so this is not directly related to dates or insufficient knowledge of the model.
But it is related to the thinking mode. (and I tested qat version)

As far as I understand, this is an attempt to protect against hallucinations, but this is the first time I have seen a model so biased that it rejects and devalues ​​any alternative information in order to stick to the original

Sign up or log in to comment