Apr 17, 02:44 PM

OLMo 1.7-7B Outperforms Llama 2-7B, 24 Point MMLU Increase, Excels on GSM8K

Allen AI has released a new version of their OLMo model, OLMo 1.7-7B, which has shown significant improvements in performance. The model scores 52 on the MMLU, surpassing Llama 2-7B and approaching the performance of Llama 2-13B, with a 24 point increase over previous versions. It also excels on GSM8K, surpassing Llama 2-13B. Key enhancements include a longer context length of 4096 tokens and the use of the new Dolma 1.7 dataset, which has contributed to the improved data quality. The team attributes these advancements to better data management, staged training, and a focus on quality in pretraining procedures.

#Allen AI #Dolma

Written with ChatGPT (GPT-4).