OLMo 2 Models Released!

>Right and totally not for looking good on benchmarks and nothing else I'm not entirely sure what you are referring to here. If you are referring to AllenAI showing in their blogpost how well their models perform on various benchmarks, I would assume that is because a garbage model would attract little attention and thus no researchers looking at or using it. It seems obvious that AllenAI would want their models to "look good on benchmarks" because of this. >There's been virtually no open model with less than 8k context for the past year, because it's useless. There have been zero fully open models released with 8k or more context that have been useful, unless I missed any? Map Neo 7b has 8k context but is almost certainly virtually useless for any practical applications. DCLM 7b and Amber 7b both have 2k context length. K2 65b has 8k context length but is much larger than the Olmo 2 models. OpenCoder 8b has 8k context but is trained mainly on coding and math. I'm also not sure how less than 8k context makes these models "useless" for performing research involving generalization, contamination, memorization and anything else that requires having full access to the model's training data. (Ideally, they would have followed LLM360's approach and uploaded model and training data checkpoints, but Olmo is still a much more open model than something like Qwen). Again, these Olmo models are the best *fully open* models, at least for their sizes. If you only care for how well a model can be run as a chatbot or code assistant or whatever, then you might as well ignore the Olmo models. There are obviously much better models to use for almost any use case *except for* ones that require having access to the model's full training data and code. I would prefer it if Meta, Mistral, Google, and all the other groups who are releasing models could be at least as open as AllenAI, but right now the Olmo models appear to be the best fully open 7b and 13b sized models available.