SmolLM2: When Smol Goes Big -- Data-Centric Training of a Small Language Model

Benchmark Model Rank Results
instruction-following-on-mtbenchSmolLM2 1.7B Instruct#15Score (1-10): 6.13
question-answering-on-natural-questionsSmolLM2 1.7B#64EM: 8.7
question-answering-on-piqaSmolLM2 1.7B#76Accuracy: 77.6
question-answering-on-piqaSmolLM2 1.7B Instruct#112Accuracy: 74.4
sentence-completion-on-hellaswagSmolLM2 1.7B#114Accuracy: 68.7
sentence-completion-on-hellaswagSmolLM2 1.7B Instruct#132Accuracy: 66.1
common-sense-reasoning-on-winograndeSmolLM2 1.7B#123Accuracy: 59.4
code-generation-on-humanevalSmolLM2 1.7B Instruct#124Pass@1: 28.1
code-generation-on-humanevalSmolLM2 1.7B#129Pass@1: 22.6
multi-task-language-understanding-on-mmlu-proSmolLM2 1.7B#134Accuracy: 19.4
multi-task-language-understanding-on-mmlu-proSmolLM2 1.7B Instruct#135Accuracy: 19.3
math-word-problem-solving-on-mathSmolLM2 1.7B Instruct#156Accuracy: 21.0
math-word-problem-solving-on-mathSmolLM2 1.7B#172Accuracy: 11.6
arithmetic-reasoning-on-gsm8kSmolLM2 1.7B Instruct#231Accuracy: 48.8
arithmetic-reasoning-on-gsm8kSmolLM2 1.7B#249Accuracy: 31.1