HKUST-DSAIL/Graph-R1-ablation-7B-RL-01reward

Model Card: Graph-R1 Series

This model card covers the Graph-R1 series of models, including the final released versions and variants used in ablation studies. All information is based on the provided research paper.

Model Details

• Model Developer: HKUST-DSAIL
• Model Series: Graph-R1
• Model Variants:
  • Graph-R1-7B: Fine-tuned from Qwen2.5-7B-Instruct-1M.
  • Graph-R1-1.5B: Fine-tuned from Qwen2.5-1.5B.
  • Ablation Models: Multiple variants based on different training configurations (e.g., data volume, training stages, reward functions, curriculum learning strategies).
• Model Type: Small reasoning language model, specialized in solving complex NP graph-theoretic problems.
• Architecture:
  • Base Model: Qwen2.5
  • Training Framework:
    1. Cold-start Supervised Fine-Tuning (SFT): Fine-tuned using long Chain-of-Thought (Long-CoT) data extracted from the QwQ-32B model to inject graph reasoning knowledge.
    2. Reasoning Optimization via Reinforcement Learning (RL): Employs a Group Relative Policy Optimization (GRPO)-based RL framework, combined with a curriculum learning strategy.
• Model Date: 2025/04

Intended Use

• Primary Use Cases:
  • Solving complex graph-theoretic computational problems at the NP-Complete level, such as the Traveling Salesman Problem (TSP), Graph Edit Distance (GED), and Maximum Clique Problem (MCP).
  • Serving as a compact, resource-efficient reasoning model for academic research and practical applications.
• Potential Cross-Domain Applications:
  • The model demonstrates transferability to other complex reasoning tasks, including mathematics, programming, STEM, and logical reasoning.