Ai2 Launches OLMo 2: New Open Source Language Models to Challenge Llama | Risks of using chatgpt in business | Top 10 most popular large language models | Openai | Turtles AI

Ai2 has unveiled the second generation of its language models, OLMo 2, which are set to compete with established models such as Meta’s Llama. Completely open source, OLMo 2 is built on an accessible dataset and offers globally competitive performance.  

Key points:

•  OLMo 2 is developed entirely with open source resources.
•  OLMo 2 models, with 7B and 13B parameters, are among the most advanced in the open landscape.
•  Training used 5 trillion tokens, including data from high-quality sources.
•  Ai2 focuses on innovation and transparency, but there is no shortage of concerns about possible abuse.

Ai2, the AI research organization founded by the late Paul Allen, has released the second version of its family of open source language models, called OLMo 2, a proposal that challenges established models such as Meta’s Llama. These new models are distinguished by being completely open, something that goes far beyond the concept of traditional open source. While models like Llama are accessible but not fully reproducible without access to all original resources, OLMo 2 meets the criteria defined by the Open Source Initiative, which states that a project should not only be accessible but also fully documented and reproducible. Ai2 has made available to the public all the tools needed to reproduce these models from scratch, including training recipes, data used and intermediate checkpoints. With OLMo 2, the goal is to promote innovation and foster equitable access to advanced technologies by making powerful and transparent resources available to the research and development communities. The new model family consists of two versions: OLMo 7B, with 7 billion parameters, and OLMo 13B, with 13 billion parameters. The parameters, in this context, represent the magnitude of the model’s computational capabilities, with the larger ones generally providing better performance in complex language processing tasks. Indeed, these models can handle complex tasks such as answering questions, writing code, summarizing texts, and performing language analysis tasks.  

As for the training process, Ai2 used a massive dataset consisting of as many as 5 trillion tokens, equivalent to hundreds of millions of words. This data was extracted from high-quality sources, including Web pages selected for their informational value, academic articles, Q&A discussion forums, and specific resources such as mathematical exercise books. The result, according to Ai2, is a set of models that are highly competitive with other open source language models, including the Llama 3.1 version of Meta. Indeed, Ai2 claims that OLMo 2 7B outperformed Llama 3.1 8B in some specific tests, thus solidifying OLMo’s position as one of the best open source language models available.  

However, the open source approach inevitably raises issues related to the security and misuse of such technologies. Although Ai2 recognizes the possibility that the models may be misused, its position is that the benefits of transparency, verifiability, and reduced concentration of power outweigh the risks. Indeed, this approach enables confidence in emerging technologies by fostering an environment in which progress is stimulated by free and universal access. Dirk Groeneveld, an engineer at Ai2, reiterated that only through full access to these models and related resources can significant ethical and accountability milestones be achieved in the field of artificial intelligence. In this way, Ai2 hopes to contribute to the creation of a more equitable and open ecosystem, where barriers to entry are broken down and research can advance for the benefit of all.  

OLMo 2 thus represents an important step forward in the evolution of open source language models, not only because of the quality of its performance, but also because of its commitment to fostering a more inclusive and accessible research environment.