according toGoogleOfficial press release, Google DeepMind The team launched a Gemma Open Model, the AI model claims to be inspired by Gemini,Available in two sizes: Gemma 2B and Gemma 7B.
▲ Image source: Google official press release (the same below)
IT Home noticed that this Gemma model can be fine-tuned using tools such as Keras 3.0 and PyTorch, and can run on low-performance devices such as laptops/IoT, and is said to have stronger performance than competing products of the same size.Said to "significantly outperform the Meta Llama-2 model" on key benchmarks.
The Google research team mentioned that this Gemma model is said to be built based on "artificial intelligence principles". In order to make the Gemma model "safe and reliable", the researchers used automated technology to filter out personal information and other sensitive data from the training set, and made a lot of fine-tuning to the model, and combined it with "reinforcement learning (RLHF)" to make the model behaviorally consistent. At the same time, the model was also subjected to "red team testing", "automatic adversarial testing" and model capability evaluation, so that the model has "responsible characteristics".