Spark-234K is a dataset of 234K scientific reasoning question-answer pairs from OpenDataArena, built from frontier research papers. It is meant for training language models on hard, self-contained science problems.
DatasetFree
Step 1: Open the dataset page on Hugging Face and review the card and the cited paper.
Step 2: Load the data with the Hugging Face datasets library using the OpenDataArena/Spark-234K identifier.
Step 3: Inspect a sample of the question-answer pairs across disciplines before training.
Step 4: Use the pairs as supervised fine-tuning data for a reasoning model.
Step 5: Cite the SPARK paper listed on the card if you use the data in published work.
It holds 234K question-answer pairs about scientific topics, generated from frontier research papers across 10 disciplines. Each item is meant to be self-contained.
Each question-answer triple is verified for self-containment, source consistency, answer coherence and reasoning grounding. A human audit of 300 samples reports 90.3% agreement with expert answers.
The pipeline includes semantic deduplication per discipline and 13-gram exact-match decontamination against downstream benchmarks.
They come from a collection of about 3.36M papers called SCI-BASE, of which about 370K frontier papers are sampled after filtering and discipline balancing.