Support Generating CoT SFT Data Using DeepSeek R1

[randydl]

Hi,

I’m exploring the possibility of using DeepSeek R1 to generate Chain-of-Thought (CoT) data for Supervised Fine-Tuning (SFT). Could we consider adding support or providing guidance on how to leverage DeepSeek R1 for this purpose?

If this is already feasible, it would be great to have some documentation or examples to help users get started. If not, I’d love to discuss the potential challenges and explore whether this could be a feature worth implementing.

Looking forward to your thoughts!

Best regards

[Stone202757]

I'm looking forward this feature also. Want to find the way to add one section like COT into the final JSON (e.g. ./augmentollkit/augmentoolkit/original/output/question_generation_generations/raw_qatuples_saved/para_0_q_0.json)

{
"paragraph": "xxx",
"metadata": "xxx",
"question": "xxx",
"answer": "xxx",
"COT": "The chain of thought for the answrer: xxxx ",
"question_group_id": "8d3b235f-b2bf-48f1-826f-593df469637c",
"paragraph_idx": 0,
"question_idx": 0
}