Overview
Logics-STEM-8B-SFT is an 8-billion parameter open reasoning model developed by Logics-MLLM, specifically engineered for Science, Technology, Engineering, and Mathematics (STEM) tasks. It is built upon the Qwen3-8B architecture and has undergone supervised fine-tuning (SFT) using the extensive Logics-STEM-SFT-Dataset-2.2M, which comprises 2.2 million long chain-of-thought examples covering a wide range of math and STEM problems.
Key Capabilities
- Specialized STEM Reasoning: Optimized for complex problem-solving in science, technology, engineering, and mathematics.
- Strong SFT Baseline: Delivers competitive performance on STEM benchmarks, often rivaling open-source RL-trained reasoners of similar scale.
- Robust Instruction Following: Maintains high accuracy in understanding and executing instructions.
- Precise Answer Formatting: Particularly adept at formatting answers for STEM multiple-choice questions, including outputting option letters within
\boxed{}. - Foundation for Further Training: Serves as an excellent starting point for advanced reinforcement learning techniques, such as RL with verifiable rewards, to further enhance reasoning accuracy and robustness.
Good For
- Developers and researchers focused on STEM-related AI applications requiring strong reasoning capabilities.
- Use cases demanding accurate and well-formatted answers for mathematical and scientific problems.
- As a base model for further fine-tuning or reinforcement learning to achieve even higher reasoning performance in specialized domains.