Hugging Face OpenCodeReasoning-Nemotron-1.1-14B Overview
OpenCodeReasoning-Nemotron-1.1-14B is a 14 billion parameter large language model developed by NVIDIA, based on the Qwen2.5-14B-Instruct architecture. It is specifically post-trained for enhanced reasoning capabilities in code generation, making it a specialized tool for competitive programming challenges.
Key Capabilities & Features
- Code Reasoning Optimization: Post-trained to excel in reasoning for code generation tasks.
- Extended Context Length: Supports a substantial context length of 64,000 tokens for processing complex coding problems.
- Strong Benchmark Performance: Achieves a Pass@1 score of 65.9 on LiveCodeBench (v5), outperforming other 14B+ distilled models and many larger models in code reasoning.
- Commercial Use: Ready for both commercial and non-commercial applications.
- NVIDIA Optimized: Designed and optimized to run efficiently on NVIDIA GPU-accelerated systems, leveraging NVIDIA's hardware and software frameworks for faster inference.
Training and Architecture
This model is a dense decoder-only Transformer. It was trained on the OpenCodeReasoning dataset, which comprises 1.165 million samples of competitive programming questions and DeepSeek-R1-0528 generated responses. The training data collection and labeling methods are hybrid (automated, human, synthetic).
Use Cases
This model is primarily intended for developers and researchers focused on building and advancing large language models, particularly those requiring strong code generation and reasoning capabilities for competitive programming or similar complex coding challenges. For more details, refer to the associated paper.