theshyustc/CoRT-Hint-Engineering-1.5B-RL

Hugging Face
TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 13, 2025Architecture:Transformer0.0K Featherless Exclusive Warm

Theshyustc/CoRT-Hint-Engineering-1.5B-RL is a 1.5 billion parameter model developed by theshyustc, based on DeepSeek-R1-Distill-Qwen-1.5B. It is trained with the CoRT (Code-integrated Reasoning within Thinking) framework, specializing in mathematical reasoning by integrating natural language with Python code execution. This model uniquely employs a Hint-Engineering approach for superior token efficiency, using 50% fewer tokens than baselines while maintaining competitive performance. It excels at complex mathematical problem-solving through strategic hint placement and multi-turn tool-integrated reasoning.

Loading preview...

CoRT-Hint-Engineering-1.5B-RL: Mathematical Reasoning with Enhanced Efficiency

CoRT-Hint-Engineering-1.5B-RL is a 1.5 billion parameter model developed by theshyustc, built upon the DeepSeek-R1-Distill-Qwen-1.5B base. It leverages the CoRT (Code-integrated Reasoning within Thinking) framework, which integrates natural language reasoning with Python code execution to tackle complex mathematical problems.

Key Capabilities & Differentiators

  • Superior Token Efficiency: Achieves competitive performance using 50% fewer tokens compared to baseline models, making it highly efficient.
  • Hint-Engineering: Employs a novel Hint-Engineering approach that strategically inserts targeted hints at critical decision points during reasoning. This optimizes code usage, balances calculation and verification (51.1% vs 48.9%), and minimizes unnecessary steps.
  • Multi-turn Tool-Integrated Reasoning: Supports interactive code execution within reasoning chains, allowing for dynamic problem-solving.
  • Strong Mathematical Performance: Demonstrates solid accuracy across various mathematical benchmarks, including AIME24 (41.0%), AMC23 (70.0%), MATH500 (85.8%), and Olympiad (55.6%), with an average accuracy of 56.4%.

When to Use This Model

This model is particularly well-suited for:

  • Mathematical Problem Solving: Excels in tasks requiring complex mathematical reasoning and precise calculations.
  • Resource-Constrained Environments: Its token efficiency makes it a strong candidate for applications where computational resources or inference costs are a concern.
  • Interactive Reasoning Systems: Designed for multi-turn, tool-integrated reasoning, making it ideal for applications that benefit from dynamic code execution and verification.

Important Note: Optimal performance requires using the specialized inference script from the CoRT GitHub repository to enable its multi-turn tool-integrated reasoning capabilities.