Google has just introduced Gemma, an innovative family of state-of-the-art open Large Language Models (LLMs), marking a significant stride in the open-source AI landscape. This release, featuring both 7B and 2B parameter models, underscores Google's ongoing commitment to open-source AI. The Hugging Face team is thrilled to support this launch, ensuring seamless integration within our ecosystem.
Gemma is available in two distinct sizes to cater to diverse computational needs: - 7B parameters: Optimized for deployment on consumer-grade GPUs and TPUs. - 2B parameters: Designed for efficient CPU and on-device applications.
Both sizes come in base and instruction-tuned variants, offering flexibility for various applications.
Our collaboration with Google has led to the best integration possible, making four open-access models (two base and two fine-tuned) readily available on the Hugging Face Hub. This integration includes: - Access to models with their respective model cards and licenses on the Hub. - Seamless integration with the 🤗 Transformers library. - Google Cloud integration for enhanced scalability. - Inference Endpoint integration for easy deployment.
Gemma stands out with its robust performance across various benchmarks, showcasing its prowess among open LLMs. The 7B variant, in particular, demonstrates remarkable capabilities, rivaling some of the best models in its class. The 2B version, while smaller, presents an attractive option for applications where size and efficiency are key considerations.
While the technical report provides insights into the training and evaluation methodologies, it leaves us curious about the specifics of the dataset composition and preprocessing. Similarly, the fine-tuning processes for the instruct models remain shrouded in mystery, piquing our interest in the underlying datasets and hyperparameters.
You can interact with the Gemma Instruct model directly on Hugging Chat, offering a hands-on experience with this advanced model. For developers looking to integrate Gemma into their projects, the 🤗 Transformers library simplifies the process, supporting a wide range of tools and utilities for both training and inference.
Created 2024-02-22T04:55:37-08:00, updated 2024-02-24T08:37:37-08:00 · History · Edit