Olmo by Ai2

Olmo by Ai2 is a fully open language model providing complete access to its model flow. The idea is to give the AI community full visibility into a state-of-the-art large language model. Olmo stands out by offering full access to its training data, architectures, and evaluation methodologies.

Olmo functions across a complete model flow, beginning with sourcing the data, which is then preprocessed, and moved onto mid-training and finally post-training. Throughout, Olmo emphasizes transparency in AI development and uses open-source tools such as OlmoCore for the training framework, Duplodocus for de-duplication, and Datamap-rs for large-scale data cleaning.

Olmo offers six different variants: 32B-Base, 32B-Think, 32B-Instruct, 7B-Base, 7B-Think, and 7B-Instruct. Each variant provides varying levels of efficacy for different areas such as programming, reading comprehension, mathematical problem-solving, and complex reasoning. Each of these variants cater to different use-cases.

In terms of programming, Olmo specifically through its 32B-Base variant achieves strong results. Its performance maintains even at extended context lengths and is effective for tool usage and RL setups.

Yes, Olmo indeed helps with reading comprehension. The 32B-Base model variant is particularly useful for achieving strong performance in reading comprehension tasks.