Curated SFT dataset for fine-tuning small language models on structured reasoning. Built from 17 open-source reasoning datasets, aggressively filtered for quality and structural integrity.

	Dataset Summary

Attribute Value

Train samples 35,944

Eval samples 734

Total size ~948 MB

Format Messages (system/user/assistant)

Languages English, Indonesian

Source datasets 17

Raw samples 202,551

Filter rate 81.6% removed… See the full description on the dataset page: https://huggingface.co/datasets/samcheng0/lumia-reasoning-sft-v1.