PKU-ML/G1-Zero-7B
PKU-ML/G1-Zero-7B is a 7.62 billion parameter causal language model developed by PKU-ML, based on the Qwen2.5-Instruct architecture, with a 32,768 token context length. It is specifically fine-tuned using Group Relative Policy Optimization (GRPO) for reinforcement learning to excel at graph reasoning tasks. The model demonstrates significant improvements on graph reasoning benchmarks like Erdos, outperforming baselines and showing strong zero-shot generalization to unseen graph tasks, while maintaining general reasoning abilities.
Loading preview...
G1-Zero-7B: Enhanced Graph Reasoning LLM
G1-Zero-7B is a 7.62 billion parameter causal language model developed by PKU-ML, built upon the Qwen2.5-Instruct architecture. It features a substantial context length of 32,768 tokens. This model is specifically trained using Group Relative Policy Optimization (GRPO) for reinforcement learning, following a supervised finetuning preliminary step, to specialize in graph reasoning tasks.
Key Capabilities
- Superior Graph Reasoning: Achieves up to 46% improvement on the Erdos benchmark compared to baselines. The 7B variant matches OpenAI's o3-mini, and the 3B model (part of the G1 series) surpasses Qwen2.5-72B-Instruct on graph tasks.
- Strong Generalization: Demonstrates zero-shot generalization to novel graph tasks, improving performance on other benchmarks like GraphWiz and GraphArena, as well as real-world graphs such as Cora and PubMed.
- Preserved General Reasoning: Maintains strong performance on general reasoning benchmarks including GSM8K, MATH, and MMLU-Pro, ensuring versatility beyond its graph specialization.
Good For
- Applications requiring advanced graph reasoning and analysis.
- Research and development in graph-based AI and machine learning.
- Tasks involving complex relational data and network structures.
- Scenarios where maintaining general language understanding alongside specialized graph capabilities is crucial.