First releases of Open-Assistant Models - February 14, 2023

Open-Assistant is a project to make an open-source chat GPT. The most recent models they have released have gone through instruction tuning and are available on the Hub. There are a variety of sizes from 1.4B to 20B parameters.

News & Announcements 📣

Hugging Face released a new library called PEFT, or Parameter-Efficient Fine-Tuning. PEFT approaches only fine-tune a small number of (extra) model parameters while freezing most parameters of the pre-trained LLMs. Check out the 🤗 PEFT: Parameter-Efficient Fine-Tuning of Billion-Scale Models on Low-Resource Hardware blog to learn more

SpeechT5 is the first Text-to-Speech model in the Transformers library, which allows you to easily create Speech Synthesis (TTS), Voice Conversion, or Automatic Speech Recognition systems.

Runway introduced Gen-1, a new model that uses language and images to generate new videos out of existing ones.

Writer open-soruced Palymra, a Language Model trained in business and marketing writing. The model comes in three sizes, from 128 million to 20 Billion parameters, available on Hugging Face.

Tutorials & Demos 📝

I wrote a blog post on how to deploy the FLAN-T5-XXL on Amazon SageMaker for inference.

Emily Webber shared how she trained Stable Diffusion on 10TB of images using Amazon SageMaker.

Moshe Wasserblat created an example of using GPT-2 for data augmentation to use smaller models with the same accuracy.

Salesforce shared a Gradio demo for BLIP-2 for an image-to-text generation.

Instructional Image Editing demo using Instruct-Pix2Pix to edit images using natural language.

Reads & Papers 📚

Meta AI introduced Toolformer, a language model that teaches itself to use various tools in a self-supervised way. The model learned to use a calculator or call an external API service.

Google Research wrote a blog post about the Flan Collection: Advancing open source methods for instruction tuning, giving insights on why the FLAN-T5 models outperform previous instruction models.

Pierre Guillou wrote a blog post about Document AI focusing on Document Understanding model at line level with LiLT, Tesseract and DocLayNet dataset.

Sebastian Raschka created a transformative reading list for better understanding Large Language Models.

Raza Habib explored if it is worth fine-tuning LLMs or if you can leverage smaller models for the same result.

The Samwald research group introduced ThoughtSource a toolchain for chain-of-thought reasoning in large language models. Checkout the repository for examples.

Multimodal Chain-of-Thought Reasoning in Language Models incorporates vision features for CoT, outperforming existing multimodal models.

Benchmarking Large Language Models for News Summarization sharing good research on how to improve abstractive summarization.

I hope you enjoyed this newsletter. 🤗 If you have any questions or are interested in collaborating, feel free to contact me on Twitter or LinkedIn.

See you next week 👋🏻👋🏻