Swahili Thinking Dataset is a collection of 20,000 ShareGPT-style chat examples for training and evaluating models that reason directly in Swahili.
Each example is a short conversation (system → user → assistant) where the assistant provides:
a final answer (messages[2].content)
a step-by-step reasoning trace in Swahili (messages[2].thinking)
