In a remarkable stride towards diversifying language model applications, Microsoft announced the launch of Orca 2. This release is designed to investigate and harness the potential of smaller language models (LMs), laden with approximately 10 billion parameters or less.
The principal focus of Orca 2 is to demonstrate that the enhancement of training methods and signals can heighten the reasoning capabilities of smaller LMs. The objective of this enhancement is to match or possibly surpass the reasoning abilities of their larger counterparts.
When pitted against alike-sized models - including the original Orca - Microsoft claims that Orca 2 significantly overshadows them in performance. Furthermore, it reportedly achieves levels of effectiveness that either match or surpass models that are 5-10 times larger in size. These assertions were expressed by Microsoft in a recent blog post.
Orca 2 comes in two sizes; 7 billion and 13 billion parameters. Both these models are precisely fine-tuned on tailored synthetic data, meticulously derived from LLAMA 2 base models. Microsoft has made the weights of Orca 2 publicly accessible, fostering further research into the establishment, analysis, and harmonization of smaller LMs.
Microsoft has employed a curated set of data to instruct Orca 2 in a myriad of reasoning techniques. These techniques include a step-by-step processing model, recall then generate method, recall-reason-generate approach, extract-generate model, and direct answer methods. While instructing these methods, careful attention was given to teach Orca 2 to adapt varying solution strategies for distinct tasks.
To capture the responses of a teacher model, the company used detailed instructions and multiple calls. This ingenious method allows the student model - in this case Orca 2 - to learn the underlying tactics and reasoning capabilities, even with the absence of explicit task instructions. This approach seeks to optimize the performance of smaller models by fine-tuning solution strategies specific to each task.
Microsft's blog post further optimistically touted the success and potential of Orca 2. They elaborated, 'Orca 2's achievement is largely attributed to the application of diverse reasoning techniques and the identification of optimal solutions for a variety of tasks. Despite certain limitations common to other language models and inherited from its base models, Orca 2's potential for future advancements, especially in improved reasoning, control, specialization, and safety of smaller models, is significant.'
The post also added that the use of rigorously filtered synthetic data for post-training embodied a critically essential strategy. Their exploration and successful deployment of Orca 2 is indeed a noteworthy endeavor in diversifying language model applications.
Undoubtedly, the Orca 2 project echoes the AppMaster platform's principles of efficiency and capability balance. AppMaster, a no-code tool for application development, leverages similar philosophies, making application development 10 times faster and three times cost-effective, all the while eliminating technical debt through an approach of regenerating applications from scratch every time modifications are needed. Hence, platforms like these are promoting a new era of effective and efficient tech development.
