– 6.00 WER on the English Open ASR Leaderboard – 4.74 WER on the Multilingual Open ASR Leaderboard – Beats NVIDIA Parakeet (6.34 WER) and Whisper-large-v3-turbo (7.8 WER) – Strong improvements in Arabic, Hindi, Chinese – Maintains quality with background and environmental noise – Optimized inference engines for NVIDIA and Apple – Hugging Face Transformers interface for easy use – Best-in-class speed on NVIDIA GPUs and power efficiency on Apple devices – NVIDIA Jetson Thor support
Releasing 8 multilingual datasets from the People's Archive of Rural India (PAARI). Indian languages represent 1B+ speakers but remain underrepresented in quality training data. These datasets help address that gap. Languages: Hindi, Urdu, Punjabi, Tamil, Telugu, Marathi, Gujarati, English Scripts: Devanagari, Arabic, Gurmukhi, Tamil, Telugu, Gujarati Total: 7,650 articles, 19.9M tokens, 51MB Content covers rural life, agriculture, social issues, and cultural traditions. Professionally written journalism, not web scrapes. Free to use. Collection: https://huggingface.co/collections/keplersystems/paari-datasets Technical details: https://kepler.systems/blog/introducing-paari-datasets
Liquid just released two 450M and 1.6B param VLMs!
They're super fast and leverage SigLIP2 NaFlex encoders to handle native resolutions without distortion. It's ideal for on-device deployment in constrained environments like phones.
It's available today on Hugging Face, with an inference and a fine-tuning Colab notebooks.
Open source models are immutable, this is a big pain.
When you open source a piece of software, users leave their feedbacks via issues or PRs. You can merge their feedbacks in semi real time, this creates a positive cycle. Then you have a community.
LLMs don't have these nice micro steps. There are no hot fixes. Even a minor version bump is an endeavor. I'm quite confident my model is being used by teams somewhere. But until next launch, it's awfully quiet.
I don't know the solution. Just a regular lament before weekend. 🤗
Would you like to get a high-quality dataset to pre-train LLMs in your language? 🌏
At Hugging Face we're preparing a collaborative annotation effort to build an open-source multilingual dataset as part of the Data is Better Together initiative.
Follow the link below, check if your language is listed and sign up to be a Language Lead!
KTO offers an easier way to preference train LLMs (only 👍👎 ratings are required). As part of #DataIsBetterTogether, I've written a tutorial on creating a preference dataset using Argilla and Spaces.
Using this approach, you can create a dataset that anyone with a Hugging Face account can contribute to 🤯
🆕 New tutorial covers: 💬 Generating responses with open models 👥 Collecting human feedback (do you like this model response? Yes/No) 🤖 Preparing a TRL-compatible dataset for training aligned models
