TianHongZXY/CHIMERA-4B-SFT
TianHongZXY/CHIMERA-4B-SFT is a 4 billion parameter language model, fine-tuned from Qwen3-4B-Thinking-2507 using supervised fine-tuning (SFT) on the CHIMERA dataset. This model specializes in enhancing performance across various benchmarks, particularly in areas like GPQA-D and HLE, demonstrating improved reasoning and problem-solving capabilities. With a context length of 32768 tokens, it is optimized for tasks requiring robust analytical understanding.
Loading preview...
CHIMERA-4B-SFT Overview
CHIMERA-4B-SFT is a 4 billion parameter language model developed by TianHongZXY. It is a supervised fine-tuned (SFT) version of the Qwen3-4B-Thinking-2507 base model, utilizing the specialized CHIMERA dataset.
Key Capabilities & Performance
This model demonstrates significant performance improvements over its base model, Qwen3-4B-Thinking-2507, primarily due to the SFT process. Notable gains are observed in:
- GPQA-D: Achieves 68.8%, an increase from 65.8%.
- AIME 24: Improves to 86.5% from 81.6%.
- HMMT Feb 25: Rises to 63.1% from 59.2%.
- HLE: Shows a substantial increase to 9.0 from 7.3.
These results indicate enhanced reasoning and problem-solving abilities, particularly in complex academic and logical tasks. The SFT alone accounts for the majority of these performance gains, with further improvements possible through additional Reinforcement Learning (RL) as seen in the CHIMERA-4B-RL variant.
When to Use This Model
CHIMERA-4B-SFT is particularly well-suited for applications requiring strong analytical performance and improved accuracy on benchmarks related to general knowledge, mathematics, and logical reasoning. Its 32768-token context length supports processing longer inputs for these complex tasks.