qihoo360/Light-IF-8B
Light-IF-8B is an 8 billion parameter large language model developed by qihoo360, specifically designed to enhance instruction following capabilities on complex tasks. It addresses 'lazy reasoning' by employing a novel framework that promotes rigorous reasoning through previewing and self-checking. The model is trained using Entropy-SFT and TEA-RL with rule-based multidimensional rewards, achieving strong performance on instruction-following benchmarks like SuperClue (0.471) and IFEval (0.932). It is optimized for scenarios requiring precise adherence to detailed instructions and complex constraints.
Loading preview...
Light-IF-8B: Enhanced Instruction Following
Light-IF-8B is an 8 billion parameter language model developed by qihoo360, engineered to significantly improve instruction following performance, particularly on complex tasks. The model addresses a common issue of 'lazy reasoning' in LLMs by introducing a unique framework that emphasizes previewing and self-checking during the thinking stage.
Key Innovations & Training
- Rigorous Reasoning Framework: Promotes planning and output verification to foster more generalizable reasoning.
- High-Quality Data Generation: Creates instruction data with complex constraints, filtered for optimal difficulty, and uses rejection sampling to build a focused dataset.
- Advanced Fine-Tuning: Utilizes Entropy-SFT (entropy-preserving supervised fine-tuning) and TEA-RL (token-wise entropy-adaptive reinforcement learning), guided by rule-based multidimensional rewards.
Performance Highlights
Light-IF-8B demonstrates notable improvements across instruction-following benchmarks:
- SuperClue: Achieves 0.471, outperforming models like Qwen3-8B (0.225) and ChatGPT-4o (0.260).
- IFEval: Scores 0.932, surpassing competitors such as Qwen3-8B (0.888) and ChatGPT-4o (0.836).
- IFBench: Achieves 0.473, competitive with top models.
Use Cases
Light-IF-8B is particularly well-suited for applications requiring:
- Precise adherence to complex instructions.
- Tasks with multiple constraints and conditions.
- Scenarios where 'thinking' and self-correction are critical for accurate output generation.