Google Introduces Gemma, a Family of Open-Source Models

Google has unveiled Gemma, a new line of compact open-source models, following the recent launches of Gemini 1.0 Ultra and Gemini 1.5 Pro. Gemma offers two variants: one with 2 billion parameters and another with 7 billion parameters.

The open-source Gemma models are commercially licensed, allowing for free usage and modification for commercial purposes, unlike the proprietary Gemini model. Despite their compact size, Gemma models are robust and built on the same research and technology as Gemini.

Running Gemma models on your laptop is effortless.

Google affirms that Gemma open-source models are compact and readily deployable on laptops or desktops. Trained on English datasets encompassing web documents, code, and mathematics, they offer versatility in various applications.

Gemma models excel in text summarization, generation, reasoning, Q&A, and beyond. Google states that Gemma models are trained on an extensive dataset totaling 6 trillion tokens.

Although the models are open-source, Google has conducted thorough testing for safety, bias, and risks. They have implemented rigorous CSAM (Child Sexual Abuse Material) filters to eliminate any harmful content and applied sensitive data filtering to exclude personal information from the models.

Google provides developers with a Responsible Generative AI Toolkit to encourage responsible use of the model. While the Gemma model family is open-source, it comes with a prohibited use policy that prohibits developers from utilizing it for “dangerous, illegal, or malicious activities,” among other restrictions.

Regarding benchmarks, the Gemma 2B model achieved a score of 42.3 in the MMLU test, while the 7B model scored 64.3. In the HellaSwag test, the 2B model achieved 71.4, and the 7B model scored 81.2. In comparison, Microsoft’s 2.7B Phi-2 model scored 56.7 in the MMLU test, and Meta’s Llama 2 (7B) scored 45.3. Google’s Gemini Nano 2 (3.2B) model scored 55.8 in the same test.

Google’s decision to release open-source models like Gemma is commendable, promoting research and fostering innovation. You can access Gemma models on Kaggle or explore the official PyTorch implementation on GitHub (visit). Additionally, you can discover Gemma on Vertex AI (visit).

Stay tuned for an in-depth test drive of this open-source model, along with comparisons to other popular open-source models like Mixtral. I’ll be providing a comprehensive, hands-on assessment of Gemma and other LLMs in the coming days.