Dataset · agentica-org
DeepScaleR Math Reasoning Pairs
agentica-org/DeepScaleR-Preview-Dataset
About 40k math problem-solution pairs from AIME, AMC, and Omni-MATH for reasoning training.
What's inside
The dataset contains roughly 40,000 unique math problem-answer pairs. Sources include AIME 1984–2023, pre-2023 AMC, Omni-MATH, and Still. Each JSON record includes the problem text and an official solution with LaTeX markup.
What you can do with it
- Building an SFT dataset for mathematical reasoning
- Fine-tuning LoRA or models to solve competition-level math problems
- Evaluating LaTeX-formatted solution generation
- Languages
- English
- Size
- 10K–100K rows
- Downloads this month
- 60.8K
- Likes
- 221
Models trained on it
- nvidia/DLER-Llama-Nemotron-8B-Merge-Research 2.4K
- nvidia/DLER-R1-1.5B-Research 2.4K
- Salesforce/E1-AceReason-14BText generation 2.3K
- mradermacher/SIRI-7B-high-i1-GGUFReinforcement learning 2.2K
- mradermacher/E1-AceReason-14B-i1-GGUF 887
- mradermacher/SIRI-7B-low-i1-GGUFReinforcement learning 823
- bartowski/junshim_When2Think-ThinkOnly-1.5B-GGUFText generation 799
- mradermacher/Segue-Qwen3_DeepScaleR-Preview-i1-GGUF 581
How to load it
from datasets import load_dataset
ds = load_dataset("agentica-org/DeepScaleR-Preview-Dataset")