New dataset type - "Thinking"

[rkinas]

Hi Maxime,
What about adding "thinking" (QwQ, O1 like) datasets?

• https://huggingface.co/datasets/PowerInfer/QWQ-LONGCOT-500K
• https://huggingface.co/datasets/PowerInfer/LONGCOT-Refine-500K
• https://huggingface.co/datasets/O1-OPEN/OpenO1-SFT-Ultra

[mlabonne]

Yes, this is my plan but some of them are not that good. I currently have this list:

• https://huggingface.co/datasets/FreedomIntelligence/medical-o1-reasoning-SFT
• https://huggingface.co/datasets/qingy2024/QwQ-LongCoT-Verified-130K
• https://huggingface.co/datasets/prithivMLmods/Deepthink-Reasoning?row=1
• https://huggingface.co/datasets/O1-OPEN/OpenO1-SFT
• https://huggingface.co/datasets/KingNish/reasoning-base-20k

[maziyarpanahi]

Please, let's only have the high-quality ones! 🙏

[rkinas]

This set looks great 👍

• https://huggingface.co/collections/Magpie-Align/magpie-reasoning-datasets-67790a13b91035bc42693885

[rkinas]

https://huggingface.co/datasets/amphora/QwQ-LongCoT-130K-2

[mlabonne]

I have an issue with Magpie in general: it's terrible for diversity. I ran ablation studies with FineTome-100k (FineLlama) vs. Magpie Ultra (UltraLlama) from Argilla and you see a significant gap. A bit of Magpie is probably okay, but it has to be mixed with more diverse samples.

[rkinas]

One more: https://huggingface.co/datasets/NovaSky-AI/Sky-T1_data_17k?row=0