Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models

Benchmark Model Rank Results
multi-task-language-understanding-on-mmluNemotron-H-56B-Base (5-shot)#9Average (%): 84.2
multi-task-language-understanding-on-mmluNemotron-H-8B-Base (5-shot)#24Average (%): 72.8
zero-shot-question-answering-on-openbookqaNemotron-H-47B-Base#9Accuracy: 48.8
zero-shot-question-answering-on-openbookqaNemotron-H-8B-Base#11Accuracy: 47.2
common-sense-reasoning-on-winograndeNemotron-H-56B-Base (5-shot)#15Accuracy: 84.5
common-sense-reasoning-on-winograndeNemotron-H-8B-Base (5-shot)#31Accuracy: 80.5
question-answering-on-piqaNemotron-H-56B-Base#18Accuracy: 85.0
question-answering-on-piqaNemotron-H-8B-Base#35Accuracy: 82.2
sentence-completion-on-hellaswagNemotron-H-56B-Base (10-shot)#19Accuracy: 89.0
sentence-completion-on-hellaswagNemotron-H-8B-Base (10-shot)#53Accuracy: 83.2
arithmetic-reasoning-on-gsm8kNemotron-H-56B-Base (8-shot)#38Accuracy: 93.7
arithmetic-reasoning-on-gsm8kNemotron-H-8B-Base (8-shot)#100Accuracy: 87.1
code-generation-on-mbppNemotron-H-56B-Base (3-shot)#38Accuracy: 77.8
code-generation-on-mbppNemotron-H-8B-Base (3-shot)#74Accuracy: 65.4
math-word-problem-solving-on-mathNemotron-H-56B-Base (4-shot)#55Accuracy: 59.4
math-word-problem-solving-on-mathNemotron-H-8B-Base (4-shot)#97Accuracy: 46.5
multi-task-language-understanding-on-mmlu-proNemotron-H-47B-Base (5-shot)#91Accuracy: 61.8
multi-task-language-understanding-on-mmlu-proNemotron-H-8B-Base (5-shot)#122Accuracy: 44.0
code-generation-on-humanevalNemotron-H-47B-Base#94Pass@1: 61
code-generation-on-humanevalNemotron-H-8B-Base#95Pass@1: 58.5