In the vast ocean of artificial intelligence, the CarperAI team at Stability AI has released two new leviathans into the wild: FreeWilly1 and FreeWilly2. These aren’t your average Large Language Models (LLMs); they’re open-source powerhouses that have been making waves in reasoning competitions across a variety of metrics.
The Birth of FreeWilly1 and FreeWilly2
FreeWilly1, the first of the duo, was fine-tuned using Supervised Fine-Tuning (SFT) in the Alpaca format, building upon the solid foundation of the original LLaMA 65B model. Not to be outdone, FreeWilly2 leveraged the LLaMA 2 70B base model to deliver performance that rivals GPT-3.5 in certain tasks.
Training the Whales
The training regimen for these models was heavily inspired by Microsoft’s innovative approach, as outlined in their article “Orca: Progressive Learning from Complex Explanation Traces of GPT-4.” The team used high-quality instructions to prompt language models, generating a dataset of 600,000 data points, which is about 10% of the dataset size used in the original Orca work.
The researchers generated 500,000 cases using a less complex LLM model and an additional 100,000 using a more complex LLM model. They meticulously screened these datasets, removing cases that originated from evaluation benchmarks to ensure valid comparisons. Their approach to synthetically generated datasets is validated by the FreeWilly models performing exceptionally well across multiple benchmarks despite training on only a tenth of the sample size used in the original Orca paper.
Evaluating the AI Models
The team used EleutherAI’s lm-eval-harness, to which they added AGIEval, to conduct evaluations of these models. The results? Both FreeWilly models are top performers when it comes to resolving complex issues in specialized fields like law and mathematics, performing intricate reasoning, and recognizing language nuances.
The Impact of the FreeWilly Models
The team believes that these two models enhance our ability to understand spoken language and unlock previously impossible possibilities. They’re excited to see all the innovative uses of these models in artificial intelligence.
For more information, check out the Reference Article and Project Page for FreeWilly1 and its successor FreeWilly2.
