• No-Sampling：No sampling is applied to the dataset, serving as the baseline.
• Phi4-Sampling：Following the Phi-4 paper, synthetic and rewritten sub-domain datasets are upsampled.

Both configs were trained with 100B tokens for performance comparison.

Evaluation

Category	Metrics (shots)	No-Sampling (100B)	Phi4-Sampling (100B)
English-Commonsense Reasoning	HellaSwag (5-shot)	0.4722	0.4827
	TruthfulQA (0-shot)	0.4114	0.4016
	Winogrande (5-shot)	0.5975	0.6117
	CommonsenseQA (5-shot)	0.2023	0.1933
	PIQA (5-shot)	0.7612	0.7644
	OpenBookQA (5-shot)	0.2640	0.3040
	BoolQ (5-shot)	0.5615	0.6621
English-Problem-Solving	ARC Easy (5-shot)	0.7222	0.7487
	ARC Challenge (5-shot)	0.3703	0.3993
	MMLU (5-shot)	0.2707	0.2652
English-Mathematics	GSM8K (5-shot)	0.0349	0.1039
	Minerva Math (4-shot)	0.0118	0.0248
Chinese	CEval (5-shot)	0.2481	0.2281
	CMMLU (5-shot)	0.2469	0.2538
Average Metrics	Average-English (w/o Math)	0.4633	0.4833
	Average-English	0.3900	0.3992
	Average-Chinese	0.2475	0.2538
	Average	0.3696	0.3888
	Average (w/o Math)	0.4274	0.4429