nthomas123/gemma4-yoruba-blindspot

Model Tested: https://huggingface.co/google/gemma-4-E2B-it

I loaded the model by going to the model’s page on Hugging Face, clicking the “Use this model” button, and then selecting Google Colab, which already provided the setup to run the model. In the model’s description, it states that it is multilingual, with a training dataset that includes content in over 140 languages. This made me curious to test whether all languages were used equally during training, especially less widely used languages compared to English.

I first tested the model by asking it a simple question in several languages, including English, Swahili, Yoruba, Bengali, Nepali, and Tagalog. I noticed that the model performed the worst when asked a question in Yoruba, so I decided to investigate its capabilities in that specific language further. The model’s responses were inconsistent, sometimes responding in Yoruba, sometimes switching to English, and often producing incoherent or repetitive output.

This brings up an even larger concern around issues of disparity and discrimination within multilingual models, which perform better in high-resource languages and poorly in low-resource languages. This reflects a broader issue of llms perpetuating technological disparity.

In order to solve this problem, one solution could be fine-tuning the model with better and higher-quality multilingual data, including low-resource languages like Yoruba. This dataset could be assembled using examples of conversation spoken in Yoruba, and if not enough real examples of Yoruba are available, a synthetic Yoruba dataset could be generated. A dataset of at least several thousand to tens of thousands of high-quality examples per language would likely be needed to improve performance.

In addition, I noticed that the model only asked to confirm the language of the input after multiple attempts. The model was not able to identify its failure to comprehend the prompt and gave inaccurate or non-sense answers. This aligns with the Chinese Room argument stating that a machine may demonstrate knowledge of a language through following guidelines and patterns while not being able to comprehend its meaning. Similarly, the model did not comprehend Yoruba but made assumptions based on its training dataset. This results in AI hallucinations when the model gives confident, yet false or meaningless answers to the prompts. This brings up a significant issue regarding the comprehension level of language models and whether their answers represent genuine understanding or pattern matching. Furthermore, it highlights a dangerous tendency among language models: giving any possible answer rather than acknowledging their inability to respond to a prompt, which could be particularly detrimental in fields such as healthcare.

Here was the code used to load the model:

Load model directly

from transformers import AutoProcessor, AutoModelForImageTextToText

processor = AutoProcessor.frompretrained("google/gemma-4-E2B-it") model = AutoModelForImageTextToText.frompretrained("google/gemma-4-E2B-it")