Dataset · qihoo360
Two-Stage SFT Dataset for Training Reasoning Models
qihoo360/Light-R1-SFTData
A collection of 79,000 examples for two-stage SFT training of models with long chains of thought.
What's inside
This is an English text dataset used by the authors from Qihoo 360 to train the Light-R1-32B model. The data is split into two JSON files for phased training: 76,000 examples in the first stage and 3,000 in the second. The dataset is designed to teach standard language models long chains of thought (long CoT) without distillation from existing reasoning models.
What you can do with it
- Training reasoning LLMs from scratch using a two-stage SFT methodology
- Fine-tuning models to solve complex AIME-level math problems
- Researching curriculum SFT for developing logical reasoning
- Tasks
- Text generation
- Size
- 10K–100K rows
- Downloads this month
- 12.4K
- Likes
- 61
Models trained on it
- mradermacher/Zhi-writing-dsr1-14b-GGUF 1.9K
- mradermacher/Zhi-Create-DSR1-14B-i1-GGUF 1.4K
- mradermacher/Zhi-writing-dsr1-14b-i1-GGUF 978
- mradermacher/Fast-Math-R1-14B-GGUF 600
- mradermacher/Zhi-Create-DSR1-14B-GGUF 388
- RabotniKuma/Fast-Math-R1-14BText generation 34
- Zhihu-ai/Zhi-Create-DSR1-14BText generation 28
- Zhihu-ai/Zhi-Create-DSR1-14B-GPTQ-INT4Text generation 18
How to load it
from datasets import load_dataset
ds = load_dataset("qihoo360/Light-R1-SFTData")