Hugging Face OpenReasoning-Nemotron-7B Overview
OpenReasoning-Nemotron-7B is a 7.6 billion parameter large language model from NVIDIA, built upon the Qwen2.5-7B architecture. It is specifically post-trained to excel in reasoning tasks related to math, code, and science solution generation, supporting an extensive context length of 131072 tokens. The model is part of a family of OpenReasoning models, available in various sizes (1.5B, 7B, 14B, 32B), all designed for high-performance reasoning.
Key Capabilities & Performance
- Specialized Reasoning: Optimized for complex problem-solving in mathematics (e.g., AIME, HMMT), coding (LiveCodeBench, SciCode), and science (GPQA, MMLU-PRO, HLE).
- Benchmark Excellence: The 7B model, along with its larger counterparts, achieves competitive scores across a suite of challenging reasoning benchmarks, often setting new records for its size class.
- Generative Solution Selection (GenSelect): Supports a "heavy" inference mode using GenSelect, a technique that combines multiple parallel generations to select the best solution. This method significantly boosts performance, particularly in math and coding, allowing the model to surpass other strong models in these domains.
- High Output Token Limit: Capable of generating solutions with up to 64,000 output tokens, crucial for detailed problem-solving.
Training & Data
The model was trained using a hybrid data collection and labeling method, incorporating questions from datasets like OpenCodeReasoning, OpenCodeReasoning-II, OpenMathReasoning, and synthetic science questions from the Llama-Nemotron-Post-Training-Dataset. All responses in the training data were generated using DeepSeek-R1-0528.
Use Cases
This model is ideal for developers and researchers focused on competitive math, code, and science problems, offering robust reasoning capabilities for generating accurate and detailed solutions.