SKU SFT-2204 · Sold by External
OpenMathInstruct-2
Product specifications
| SKU | SFT-2204 |
|---|---|
| Data type | SFT dataset |
| Volume | 14M problem-solution pairs |
| Size on disk | Not published by source |
| Format | Parquet (Hugging Face) |
| Access model | PUBLIC LICENSE |
| Pricing | Free · open-source license |
| Quality score | — |
| License | CC-BY-4.0 |
OpenMathInstruct-2 is NVIDIA's large-scale math SFT corpus: 14M problem-solution pairs in the train split, produced by augmenting GSM8K and MATH problems and generating solutions with Llama-3.1-405B-Instruct. Solutions are text chain-of-thought (final answer summarized); unlike OpenMathInstruct-1 (~1.8M pairs, which mixed text reasoning with executed Python code-interpreter blocks), OMI-2 is text CoT. Downsampled train_1M / train_2M / train_5M subsets are provided for smaller budgets.