StabilityAI Introduced Two Insane New AI Models: FreeWilly1 and FreeWilly2

Stability AI has launched two new AI models called Free Willy 1 and Free Willy 2, and I have to say both of them are truly impressive. Now let's get to know what these models are about. So both are based on Llama foundation models created by Meta. LLAMA is short for Large Language Model with Adaptive Learning, and it uses a unique way of training that adjusts itself based on the kind of data it receives and the job it needs to do. Because of this, LLAMA models are very powerful and versatile and they can handle a wide range of natural language tasks such as text generation, summarization, question answering, sentiment analysis, and more. Now looking at FreeWilly1, it uses the Llama 65B model which has 65 billion parameters. On the other hand, FreeWilly2 uses the Llama 270B model which has 70 billion parameters. This 70B model is like an updated version of the 65B model with better performance and efficiency. Both Free Willy models were improved using a method called Supervised Fine Tune, or SFT for short. This method involves giving the models detailed instructions to help them learn. These instructions are written in natural language and describe what the model should do or output given a certain input or context. For instance, an instruction might say, sum up this article in three sentences or write a short poem about love. By using sft, Free Willy models can learn to perform complex tasks that require reasoning and understanding of natural language nuances. The training methodology for Free Willy models was inspired by a groundbreaking approach from Microsoft research described in their paper ORCA Progressive Learning from Complex explanation traces of GPT4. In this paper, ORCA Microsoft researchers showed how they trained a small model called ORCA by imitating the outputs and explanations of a large foundation model like GPT4, which is one of the most advanced language models in the world with 175 billion parameters and incredible capabilities. However, it is not publicly available and it is very expensive and difficult to train and run. What Microsoft researchers did was they used GPT4 to generate synthetic data for Orca by prompting it with various questions and tasks. For each input, GPT4 would not only produce an output, but also an explanation trace that shows how it arrived at that output step by step. For example, if the input was what is the capital of France? GPT4 would output Paris, but also explain how it used its knowledge base, its memory, its logic, and its common sense to answer the question. ORCA would then learn from these explanation traces by imitating GPT4's reasoning process and trying to reproduce its outputs. Stability AI followed a similar approach for Free Willy models, but they used different sources of data and instructions. Instead of using GPT4 as their teacher model, they used ChatGPT. They also used different datasets created by Enrico Schiphol, a researcher who specializes in creating high quality instructions for language models. These datasets include Caotsubmix Original, NIV2 Submix Original, Flan 2021 Submix Original and T0 Submix Original. These datasets contain various natural language tasks such as text classification, generation, summarization, translation, analysis, paraphrasing and more. Using these datasets, Stability AI generated 600,000 data points for free willy models by prompting ChatGPT with high quality instructions and collecting its outputs and explanations, which is about 10% of the data size used by Microsoft for Orca and that shows how efficient and effective this method is. They also carefully filtered the data to remove any examples that came from evaluation benchmarks to ensure fair comparisons and avoid overfitting. Now Stability AI evaluated Free Willy models using various benchmarks that measure their natural language, understanding and reasoning abilities. These benchmarks include Open LLM Leaderboard, GPT for all AGI eval and professional and academic exams such as sat, lsat, GRE and gmat. The results show that these models outperform many state of the art instruction tuned models such as Vicuna 13B, Bard and text DAVINCI003. They also reach parity or surpass chatgpt on some tasks and even come close to gpt4 on others. Lets look at some examples. The Open LLM Leaderboard is a test of how well a model can adapt to different language tasks. Free Willy 2 got 103 points on this test which is better than ChatGPT's 100 points and Vicuna 13B's 76 points. The GPT for all test looks at how well a model can do on 20 different language tasks without any special training. Here Free Willy 2 got 47 points, beating out Vicuna 13B's 30 points and and Text DaVinci 003's 42 points. The AGI Eval test uses big academic and professional tests to see how well a model can solve problems. On this one Free Willy 2 got 45 points. That's just a bit less than ChatGPT's 49 points, but way better than Vicuna 13B's 20 points. Free Willy 2 also did really well on some tough tasks. On the heliswag task where the model has to figure out the best ending for a story, the Free Willy 2 was right 86.4% of the time. That's better than ChatGPT, which was right 85.5% of the time. On the Winogrande task, the model needs to figure out what a pronoun like he or she is referring to in a sentence. Free Willy 2 got it right 79.8% of the time, while ChatGPT was right 78.9% of the time. Finally, on the SAT math task, the model has to solve math problems like the ones on the sat test. Free Willy 2 got 63.6% of them right. That's a bit less than ChatGPT's 65.5%, but it's still really good for a model that's all about language. To ensure the validity and reliability of these results, Stability AI used two different tools to conduct their evaluations, Eleutherai's LM Eval Harness and Hugging Face's Open LLM Leaderboard. Eleutherai's LM Eval Harness is a tool that allows researchers to evaluate their language models on various natural language tasks using standardized metrics and protocols. Stability AI added AGI eval to this tool to include more challenging tasks that test the model's reasoning and problem solving skills. Hugging Faces Open LLM Leaderboard is a platform that allows researchers to submit their language models and compare their performance with other models on various natural language tasks using using standardized metrics and protocols. Stability AI submitted their Free Willy models to both tools and verified that their results were consistent and reproducible. They also invited Hugging Face to independently reproduce their results on July 21, 2023 and publish them on their leaderboard. Hugging Face confirmed that they were able to reproduce stability AI's results with minimal differences due to random seeds and hardware variations. The spokesperson for Stability AI and a researcher behind the Free Willy models, Enel Islamovich, expressed pride in the Free Willy models, believing they will impact the open source LLM community. He highlighted their power, affordability and accessibility, emphasizing their potential in advancing learning from explanations by humans or AI. Islamovic is enthusiastic about the model's potential in natural language understanding and reasoning. They could tackle challenges in natural language processing such as common sense reasoning and open doors for novel applications like interactive storytelling and educational content creation. However, there were hurdles during their development. Ensuring quality and diverse data was paramount with a need to avoid biases and duplications. The team grappled with balancing model size and operational speed since bigger models, though powerful, can be costly and slow. Islamovic admitted the models aren't flawless. Their reliance on Chatgpt which isn't as advanced as GPT4, can be a drawback. Inaccuracies in ChatGPT can influence free Willy's learning, especially when faced with unfamiliar or ambiguous inputs. He explained that his team emphasized safety and ethics when developing their models. They adhered to responsible AI practices like transparency and fairness, and conducted thorough tests to ensure the models were free from issues like bias and misinformation. He welcomed input and collaboration from anyone in the AI community to improve the models and advance AI for the good of all. I hope you enjoyed this video and learned something new today. If you did, please give it a thumbs up and subscribe to my channel for more AI content. Also, let me know in the comments what you think about these models. Thank you so much for watching and I'll see you in the next video. Sam.