Google has released three new additions to its Gemma family of lightweight, state-of-the-art open models: Gemma 2 2B, ShieldGemma, and Gemma Scope. Gemma 2 2B is a 2.6B parameter version, making it suitable for on-device use. ShieldGemma is a series of safety classifiers trained on top of Gemma 2, while Gemma Scope is a comprehensive suite of sparse autoencoders for Gemma 2 2B and 9B.
AIGoogle | Rating: 62 | 2024-07-31 04:19:36 PM |
Researchers have published a new approach to memory-efficient diffusion transformers with quantization, which aims to address the increasing memory requirements of large-scale transformer-based diffusion backbones for text-to-image generation. The new method uses quantization to reduce the memory footprint of the models, making them more scalable and efficient. The update is available on GitHub, and the research is focused on improving the performance of high-resolution text-to-image generation models.
AIHuggingFace | Rating: 62 | 2024-07-30 01:39:27 AM |
Hugging Face and NVIDIA have launched a new service on the Hugging Face Hub, allowing Enterprise Hub organizations to use open models with NVIDIA's accelerated compute platform, DGX Cloud, for serverless inference. This service enables users to run inference on popular Generative AI models, including Llama and Mistral, using standardized APIs and a few lines of code. The collaboration simplifies access and use of NVIDIA AI technology.
AINVIDIA | Rating: 62 | 2024-07-29 08:29:44 PM |
Researchers developed Docmatix and noticed that fine-tuning Florence-2 on it resulted in low scores on the benchmark. They had to fine-tune the model further on DocVQA to learn the syntax required for the benchmark. The generated answers semantically align with the reference answers but still receive low scores, raising questions about whether fine-tuning is necessary.
AIHuggingFace | Rating: 57 | 2024-07-25 04:19:28 PM |
Hugging Face has released Llama 3.1, a new iteration of the Llama family, in collaboration with Meta. The new models come in three sizes: 8B, 70B, and 405B, with base and instruction-tuned variants. The models are available on the Hub, with eight open-weight models, including three base models and five fine-tuned ones.
AIHuggingFace | Rating: 71 | 2024-07-23 03:16:19 PM |
Apple unveiled Apple Intelligence at WWDC 24, demonstrating its commitment to efficient, private, and on-device AI. The technology powers AI-enhanced features in various apps and the OS, which are integrated with Apple Silicon's capabilities. The company reiterated its focus on practical uses for everyday tasks, showcasing time-saving and fun helpers.
AIApple | Rating: 57 | 2024-07-22 04:56:11 PM |
Docmatix, a huge dataset for Document Visual Question Answering, has been published. It is 100s of times larger than previously available datasets. The dataset was created to address a gap in the availability of large-scale Document Visual Question Answering datasets. Fine-tuning Florence-2 using this dataset shows a 20% increase in performance on DocVQA.
AIHuggingFace | Rating: 73 | 2024-07-18 02:56:27 PM |
TGI has introduced Multi-LoRA, a solution that allows organizations to deploy once and serve 30 models. This technology aims to simplify the process of managing multiple AI models, which are often built for specific tasks. The solution is motivated by the fact that smaller, specialized models can outperform larger, general-purpose models on specific tasks.
AIHuggingFace | Rating: 62 | 2024-07-18 02:36:20 PM |
Microsoft has introduced SmolLM, a family of small language models with 135M, 360M, and 1.7B parameters, trained on a new high-quality dataset. These models are designed to operate on local devices, reducing inference costs and improving user privacy. The models are part of a trend in small language models, which involve techniques such as distillation or quantization to compress large models or training small models from scratch on large datasets.
AIMicrosoft | Rating: 62 | 2024-07-16 04:36:23 PM |
The article discusses how to create an Argilla 2.0 Chatbot that can understand technical documentation and chat with users about it. The process involves creating a synthetic dataset from the technical documentation to fine-tune a domain-specific embedding model. The Chatbot is then deployed to a Hugging Face Space, allowing users to interact with it and store the interactions.
AI | Rating: 56 | 2024-07-16 07:16:16 AM |