Trained on a mixture of 2.1k examples, most of which were manually gathered and verified.
The model will think much shorter than QWQ and R1 models thanks to brief but high quality SFT dataset.
System messages to trigger think tags (which should be placed under "<start_of_turn>user"):
"You are an expert assistant. Think using <think> tags."
"You are a thinking assistant."
Example:
<start_of_turn>user
You are a thinking assistant.
(1,234)² means 1,234 * 1,234; (1,234)³ means 1,234 * 1,234 * 1,234; and so forth. When (1,234)²³ is completely multiplied out, what will the number be in the ones place?
...
The model can miss the <think> tag sometimes. You can add this to template, or try again with a different seed.
The model can also think about an image as well:
