s1K-1.1
Just 1,000 ultra-curated questions with reasoning traces — selected from 59k candidates for difficulty, diversity, and quality, with traces regenerated by DeepSeek R1 in the 1.1 release. Fine-tuning Qwen2.5-32B on s1K (26 minutes on 16 H100s) plus 'budget forcing' at inference exceeded o1-preview on competition math. The LIMA of reasoning: proof that data quality can beat quantity.
from datasets import load_dataset
ds = load_dataset("simplescaling/s1K-1.1")Preview a sample row
{
"question": "How many positive integers n ≤ 1000 are divisible by neither 5 nor 7?",
"solution": "686",
"cot_type": "math",
"deepseek_thinking_trajectory": "By inclusion-exclusion: 1000 - 200 - 142 + 28 = 686..."
}Fine-tune with this dataset
Estimated VRAM to fine-tune with QLoRA (4-bit base model + LoRA adapters), using conservative defaults:
New to fine-tuning? Follow the step-by-step walkthrough: Fine-Tune Your First LLM in 1 Hour
Frequently asked questions
Can I use s1K-1.1 commercially?
Yes — s1K-1.1 is released under Apache 2.0, a permissive license that allows commercial use, including training models you ship in a product. Check the dataset card for attribution requirements before release.
How much data does s1K-1.1 contain, and do I need all of it?
s1K-1.1 contains 1k Rows. You rarely need all of it: for style and format fine-tuning, a few hundred to a few thousand examples are enough — load a slice (e.g. split="train[:1000]") and scale up only if quality plateaus.
What is s1K-1.1 best used for?
Cheap, fast reasoning fine-tunes — 1k samples means minutes of training, not days. It belongs to the Reasoning section of our dataset hub, where you'll find alternatives and complementary sets.