C4AI released the Aya Expanse family of multilingual AI models, addressing the performance gap between multilingual and monolingual AI. These models, with 8B and 32B parameters, were developed using data arbitrage, multilingual preference training, safety tuning, and model merging techniques.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 80 | 2024-10-24 01:12:06 PM |
Google DeepMind and Hugging Face launched SynthID Text, a technology that allows for watermarking AI-generated text and detecting those watermarks. This aims to increase trust in information and address issues like misattribution and misinformation.
Source: https://huggingface.co/blog
AIGoogle | Rating: 86 | 2024-10-23 05:32:21 PM |
HUGS is launched to simplify and accelerate AI application development with open models, using Hugging Face technologies and optimized for various hardware accelerators.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 80 | 2024-10-23 04:02:18 PM |
CinePile 2.0, an improved version of a long video QA dataset, has been released. It features a new approach called adversarial dataset refinement, which aims to strengthen existing datasets and be part of future dataset creation pipelines. CinePile was initially launched in May 2024 with approximately 300,000 training samples.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 86 | 2024-10-23 03:02:16 PM |
On October 22, 2024, outlines-core 0.1.0 was released with structured generation in Rust and Python. This update brings a 2x improvement in index compilation speed, easier incorporation into other libraries due to separation of concerns, and increased portability with Rust's core algorithms allowing bindings for languages other than Python. This release is a result of collaboration between dottxt and Hugging Face.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 86 | 2024-10-22 05:42:35 PM |
Stable Diffusion 3.5 Large, an improved version of Stable Diffusion 3, is now available on the Hugging Face Hub and can be used with the 🧨 Diffusers library. It features an 8B parameter model and a timestep-distilled model for faster inference.
Source: https://huggingface.co/blog
2024-10-22 03:22:05 PM |
Transformers.js v3 has been released with new features including WebGPU support for high-performance computations, updated models, and additional tasks. It can be installed via NPM or a CDN.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 80 | 2024-10-22 02:52:25 PM |
Hugging Face launched Speech-to-Speech (S2S), a project that uses a pipeline of models to convert speech to synthesized speech. The pipeline includes Voice Activity Detection, Speech to Text, a Language Model, and Text to Speech. S2S supports multiple languages including English, French, Spanish, Chinese, Japanese, and Korean.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 92 | 2024-10-22 03:30:40 AM |
Llama 3.2 is now available in Keras, allowing users to load and generate text using the model from standard Hugging Face checkpoints.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 86 | 2024-10-21 02:43:15 PM |
On October 16, 2024, a GitHub update from Unsloth addressed an issue with gradient accumulation in the transformers Trainer. The problem, initially reported by @bnjmn_marie, is that losses don't match between training runs when gradient accumulation is turned on and off. The issue seems to be in the modeling code of each model, where a default loss function may not be correctly calculating the loss for tasks like question answering, token classification, causal LM, or masked LM.
Source: https://huggingface.co/blog
AIHuggingFace | Rating: 90 | 2024-10-16 08:16:49 PM |
