qihoo360/Light-IF-14B
Light-IF-14B is a 14 billion parameter instruction-following language model developed by Qihoo360, featuring a 32768-token context length. It achieves state-of-the-art performance on complex Chinese instruction following benchmarks like SuperCLUE-CPIF, outperforming larger models. This model is specifically optimized for accurately following multi-constraint instructions through a carefully designed curriculum learning strategy.
Loading preview...
Model Overview
Light-IF-14B, developed by Qihoo360, is a 14 billion parameter instruction-following model with a 32768-token context length. It is noted for its exceptional performance in complex instruction following, particularly in Chinese, where it has achieved domestic SOTA on the SuperCLUE-CPIF benchmark, surpassing models like ERNIE-X1.1 and DeepSeek-V3.2-Exp-Thinking.
Key Capabilities
- Advanced Instruction Following: Excels at accurately interpreting and executing complex, multi-constraint instructions.
- Curriculum Learning: Its strong performance is attributed to a sophisticated curriculum learning strategy, progressively increasing instruction difficulty during SFT and two-stage reinforcement learning.
- Benchmark Performance: Demonstrates leading results across various instruction-following evaluations, including SuperCLUE, IFEval, CFBench, and IFBench, often outperforming larger and more established models.
Use Cases
Light-IF-14B is particularly well-suited for applications requiring precise adherence to detailed and intricate instructions, especially in Chinese language contexts. Its ability to handle complex prompts makes it valuable for tasks where exact output formatting and constraint satisfaction are critical.