Hugging Face released MTEB (Massive Text Embedding Benchmark), a text embedding model evaluation benchmark covering 56 datasets, 8 task categories, and supporting up to 112 languages; the accompanying paper introduces task design and result analysis, while the GitHub repository provides evaluation code and model submission procedures; users can quickly evaluate any embedding model via pip install mteb and submit results to the public leaderboard.
Hugging Face has launched the Inference Endpoints service, enabling users to deploy machine learning models from Hugging Face Hub to managed infrastructure on cloud platforms such as AWS with a single click.
Starting from version 0.5.1, Hugging Face Diffusers supports the JAX/Flax implementation of Stable Diffusion, enabling high-speed inference on TPUs in Colab, Kaggle, or Google Cloud Platform.
The Hugging Face blog provides a detailed retrospective on building an efficient inference service for the open-source large model BLOOM (176B parameters): from porting the model to the transformers library, to adopting Tensor Parallelism instead of Pipeline Parallelism.
Hugging Face announced the availability of Digital Object Identifier (DOI) generation for models and datasets on the Hub. Users can request a DOI directly in repository settings, supported through collaboration with DataCite. The DOI is linked to metadata and updates with new versions, while older versions automatically expire. A one-click citation button is provided on the page.
rinna has released Japanese Stable Diffusion, a text-to-image model fine-tuned from Stable Diffusion specifically for Japanese. It supports Japanese prompt input and generates images reflecting Japanese culture. The model is open-sourced on Hugging Face and GitHub. Training employs a two-stage process: the first stage replaces the text encoder using a Japanese SentencePiece tokenizer, and the second stage jointly fine-tunes the text encoder and the latent diffusion model.
Hugging Face has launched a zero-shot evaluation feature on the Hub, enabling code-free zero-shot text classification evaluation for any causal language model (up to 66 billion parameters), implemented via AutoTrain; using WinoBias as an example, it demonstrates an inverse scaling trend in gender bias tasks as model size increases.
DALL·E is now available without a waitlist, enabling new users to begin creating immediately; OpenAI stated that accumulated deployment experience and improvements to the safety system made broader access possible.
Hugging Face AutoTrain has added support for image classification tasks. Users can upload data (local folders or datasets on the Hub such as NimaBoscarino/butterflies) without code or configuration, select automatic model search, train up to 5 candidate models (free if there are fewer than 500 images and 5 candidate models are selected), ultimately obtain a model with up to 84% accuracy, and try it directly via the Hub inference widget.
A Hugging Face blog post details how the Accelerate library leverages PyTorch's meta device to create empty models and automatically infers a device_map to distribute weights across devices (GPU/CPU/disk).
Hugging Face, in collaboration with Intel Labs and UKP Lab, has released the SetFit framework, which supports few-shot fine-tuning of Sentence Transformers.
OpenAI has trained and open-sourced a neural network named Whisper, which achieves near-human-level robustness and accuracy in English speech recognition tasks.
The official Hugging Face blog introduces three open-source solutions for achieving ultra-fast inference of the BLOOM-176B model on 8x80GB A100 GPUs: Accelerate (supporting bf16/int8.
Hugging Face's official blog announced the release of version v0.3 of the diffusers library, adding image-to-image generation, Textual Inversion, and experimental inpainting.
The Hugging Face blog published a tutorial guiding readers on how to train an offline decision transformer model from scratch using the transformers library and a custom Data Collator on Google Colab.
A Hugging Face blog post published a tutorial guiding users on how to train language models on GPUs using the Megatron-LM framework developed by NVIDIA, and convert them into 🤗 Transformers compatible formats.
The official Hugging Face blog published a tutorial detailing how to run the Stable Diffusion v1-4 model using the 🧨 Diffusers library, including installing dependencies.
Hugging Face, in collaboration with BigScience, introduced the LLM.int8() method, enabling lossless 8-bit quantized inference for ultra-large language models such as BLOOM-176B.
Hugging Face has released a new library, skops, designed to host scikit-learn models on the Hugging Face Hub, automatically generate compliant model cards (including YAML metadata and Markdown content), enable inference widgets, and support environment configuration, tabular evaluation metrics, and visualization embeddings such as confusion matrices; example code demonstrates the complete workflow from training a DecisionTreeClassifier to pushing the repository.
Hugging Face has launched Private Hub, a private ML collaboration platform deployable in customers' own compliant environments. It supports three deployment methods: on-premises, on-cloud, and managed SaaS. Features include AutoTrain for automatic fine-tuning, Inference API for rapid deployment, Spaces for demo applications, private hosting of models/datasets, and SSO access control.
The official Hugging Face blog introduces how to accelerate text generation in the TensorFlow implementation of the 🤗 transformers library via XLA compilation. Tests show speedups of up to 100x, making it faster than PyTorch. The article provides complete code examples, padding adaptation techniques, usage of jit_compile, and performance benchmark comparisons.
DALL·E is now open for beta testing; OpenAI will invite 1 million people from the waitlist over the next few weeks. Users can generate images using a monthly auto-replenished free quota, or purchase additional credits in bundles of 115 generations for $15.
The Hugging Face blog published a tutorial introducing the Dynamic Adversarial Data Collection (DADC) method. Using the MNIST handwritten digit recognition task as an example, it demonstrates how humans can write adversarial samples on the 🤗 Spaces canvas, label incorrect predictions, and iteratively train the model with new samples. The article includes PyTorch model definitions, Gradio interaction implementations, and complete Space links.