YueLinHu/GAC-Qwen3.5-4B
YueLinHu/GAC-Qwen3.5-4B is a 4.5 billion parameter language model based on the Qwen3.5-4B backbone, enhanced with GAC noise-aware adaptive SFT–RL post-training and supervised mathematics and general-dialogue refinement. It features a 32768 token context length and is optimized for reasoning tasks, particularly in mathematics and logic, as well as code generation. This model is designed for research, reasoning experiments, and code-generation evaluation, offering strong performance across various benchmarks including AMC, HumanEval, and BBH.
Loading preview...
GAC-Qwen3.5-4B Overview
GAC-Qwen3.5-4B is a 4.5 billion parameter language model built upon the Qwen3.5-4B backbone, distinguished by its unique post-training methodology. It incorporates GAC noise-aware adaptive SFT–RL (Supervised Fine-Tuning and Reinforcement Learning) combined with a full-parameter supervised refinement stage using specialized mathematics and general-dialogue examples. This hybrid approach aims to enhance the model's reasoning capabilities and overall performance.
Key Capabilities
- Advanced Reasoning: Demonstrates strong performance in complex reasoning tasks, particularly in mathematics (e.g., AMC, AIME) and logical deduction (BBH).
- Code Generation: Achieves competitive results in code generation benchmarks like HumanEval (86.6% pass@1) and MBPP (68.9% pass@1).
- Robust Evaluation: Released with a comprehensive three-seed evaluation summary across 11 task slices, providing insights into decoding variability and performance consistency.
- High Context Length: Supports a substantial context window of 32,768 tokens, facilitating the processing of longer inputs and complex problem descriptions.
Good For
- Research and Development: Ideal for researchers exploring advanced post-training techniques and their impact on model performance.
- Reasoning Experiments: Suitable for tasks requiring strong mathematical, scientific, and logical reasoning.
- Code Generation Evaluation: A valuable tool for evaluating and developing code generation applications, especially in competitive programming or problem-solving contexts.
- Text-Only Applications: While retaining a text-and-vision architecture, its evaluations focus on text-only reasoning and code, making it well-suited for these specific domains.