by NVIDIA
Nemotron Math Proofs v3 SFT is an NVIDIA dataset of long-form proof generation, refinement, verification and meta-verification traces. It is for training language models on mathematical reasoning.
DatasetFree
Step 1: Open the dataset page on Hugging Face and review the README.
Step 2: Load the train split with the Hugging Face datasets library, which reads the 10 JSONL shards as one split.
Step 3: Filter rows by metadata.data_type to pick proof, refinement, verification or meta_verification samples.
Step 4: Use the messages field as supervised fine-tuning data for your model.
Step 5: Credit the dataset according to the CC BY 4.0 license.
CC BY 4.0. NVIDIA says the dataset is ready for commercial or non-commercial use.
Use metadata.data_type, which is proof, refinement, verification or meta_verification. The subset field is only kept for compatibility.
It has 414,890 samples from 15,818 unique problems, with a total disk size of about 122 GiB.
They are selected from the AoPS subset of an earlier Nemotron proofs dataset, using hard problems based on a prior pass-rate evaluation.