Knowledge Distillation During Mid-Training Favors Reasoning over Factual Recall

Benchmark Model Rank Results
question-answering-on-natural-questionsSwitch Distillation (1B; 7B teacher, q30% mid-training)#51EM: 25.3
multi-task-language-understanding-on-mmlu-proSwitch Distillation (1B; 7B teacher, after DPO)#132Accuracy: 20.1
common-sense-reasoning-on-winograndeSwitch Distillation (1B; 7B teacher, after SFT)#142Accuracy: 56.1
math-word-problem-solving-on-mathSwitch Distillation (1B; 7B teacher, full post-training)#155Accuracy: 22.4
arithmetic-reasoning-on-gsm8kSwitch Distillation (1B; 7B teacher, full post-training)#157Accuracy: 79.8