Oral
in
Workshop: The Fourth Workshop on Efficient Natural Language and Speech Processing (ENLSP-IV): Highlighting New Architectures for Future Foundation Models
OLMoE: Open Mixture-of-Experts Language Models
Niklas Muennighoff · Luca Soldaini · Dirk Groeneveld · Kyle Lo · Jacob Morrison · Sewon Min · Weijia Shi · Evan Walsh · Oyvind Tafjord · Nathan Lambert · Yuling Gu · Shane Arora · Akshita Bhagia · Dustin Schwenk · David Wadden · Alexander Wettig · Binyuan Hui · Tim Dettmers · Douwe Kiela · Noah Smith · Pang Wei Koh · Amanpreet Singh · Hannaneh Hajishirzi
Keywords: [ Efficient Training ]
We introduce OLMoE, a fully open, state-of-the-art language model leveraging sparse Mixture-of-Experts (MoE). OLMoE-1B-7B has 7 billion (B) parameters but uses only 1B per input token. We pretrain it on 5 trillion tokens and further adapt it to create OLMoE-1B-7B-Instruct. Our models outperform all available models with similar active parameters, even surpassing larger ones like Llama2- 13B-Chat and DeepSeekMoE-16B. We present novel findings on MoE training, define and analyze new routing properties showing high specialization in our model, and open-source all our work: model weights, training data, code, and logs.