With recent releases like ChatGPT, and Bard, the user experience is taking centre stage where chatbot development is booming.
Open access and open-source alternatives are driving interest in the chatbot space.
Despite the significant efforts to create frameworks for training chatbots, such as trlX, trl, DeepSpeed Chat, and ColossalAI, there remains a scarcity of open access and open source models incorporating both instruction finetuning and reinforcement learning through human feedback (RLHF) paradigms.
As Open Assistant, Anthropic, and Stanford have begun to release chat RLHF datasets to the public, the time is ripe for the emergence of a large-scale instruction fine-tuned RLHF model.
Stability AI has answered the call with the launch of StableVicuna, the first open-source RLHF LLM chatbot of its kind.
StableVicuna: A Game-Changer in the Chatbot Arena
Built on Vicuna v0 13b, an instruction fine-tuned LLaMA 13b model, StableVicuna combines Vicuna as the base model with a three-stage RLHF pipeline.
This process, outlined by Steinnon et al. and Ouyang et al., involves further training the base model using supervised finetuning (SFT) with three datasets: OpenAssistant Conversations Dataset (OASST1), GPT4All Prompt Generations, and Alpaca.
With the use of trlX, a reward model is trained on RLHF preference datasets such as OASST1, Anthropic HH-RLHF, and Stanford Human Preferences (SHP).
Finally, Proximal Policy Optimisation (PPO) reinforcement learning is performed to complete the RLHF training, resulting in the groundbreaking StableVicuna chatbot.
How to Get Your Hands on StableVicuna-13B
Available on the HuggingFace Hub, StableVicuna can be downloaded as a weight delta against the original LLaMA model.
To obtain StableVicuna-13B, users need to download the weight delta and apply for LLaMA weights separately. Once both the weight delta and LLaMA weights are obtained, a script provided in the GitHub repo can be used to combine them and create StableVicuna-13B.
A Sneak Peek at the Upcoming Chatbot Interface
Stability AI is also giving users a glimpse of their upcoming chat interface, which is in the final stages of development. The interface promises a seamless user experience, further elevating the value of StableVicuna in the chatbot market.
The release of StableVicuna is only the beginning. Stability AI plans to iterate on the chatbot and deploy a Discord bot to the Stable Foundation server in the coming weeks. Users are encouraged to try StableVicuna and provide feedback to help improve the user experience.
Introducing DeepFloyd IF: A Powerful Text-to-Image Model
In addition to StableVicuna, Stability AI and its multimodal AI research lab DeepFloyd, announced the research release of DeepFloyd IF, a cutting-edge text-to-image cascaded pixel diffusion model.
With an impressive zero-shot FID score of 6.66 on the COCO dataset and advanced features like deep text prompt understanding, application of text description into images, and zero-shot image-to-image translations,
DeepFloyd IF is poised to make a significant impact in the AI industry.
DeepFloyd IF was trained on a custom high-quality LAION-A dataset that contains 1 billion (image, text) pairs.
LAION-A, an aesthetic subset of the English part of the LAION-5B dataset, was refined through deduplication, extra cleaning, and other modifications. DeepFloyd’s custom filters removed watermarked, NSFW, and other inappropriate content.
Licensing and the Future of DeepFloyd IF
Initially, DeepFloyd IF is being released under a research license, with plans to move to a permissive license release after incorporating user feedback.
The research on DeepFloyd IF has the potential to unlock novel applications across various domains, including art, design, storytelling, virtual reality, accessibility, and more.
To fully leverage all available functionalities of this state-of-the-art text-to-image model, researchers can create innovative solutions that benefit a wide range of users and industries.
Pioneering the Future of Chatbots and AI Models
With the release of StableVicuna and DeepFloyd IF, Stability AI is proving itself to be a driving force in the world of chatbots and AI models.
The company’s commitment to continuous improvement and user feedback indicates a dedication to providing the best possible user experience.
As a leading voice in the AI community, Stability AI is focused on developing innovative solutions that have far-reaching applications and benefits.
The success of StableVicuna and DeepFloyd IF is a testament to the company’s ability to adapt and evolve in the ever-changing landscape of artificial intelligence.
