youngzhong/SOD-1.7B

SOD-1.7B: A Distilled Agentic Language Model

SOD-1.7B is a 1.7 billion parameter student model developed by YoungZhong, distilled from a 4 billion parameter teacher model using a novel technique called Step-wise On-policy Distillation (SOD). This method is specifically designed to train small language model agents with robust tool-integrated reasoning capabilities.

Key Differentiators & Capabilities

• Addresses Cascading Error Propagation: SOD introduces an adaptive step-level weighting mechanism to suppress distillation loss on drifted steps and restore supervision when the student model recovers alignment, all with negligible additional computational cost.
• High Performance in Agentic Reasoning: The model demonstrates strong performance on challenging math, science, and code benchmarks (AIME, GPQA-Diamond, LiveCodeBench-v6).
• Efficient Knowledge Transfer: It recovers 69.8% of its 4B teacher's performance with only 1.7B parameters, achieving an average of 42.98% across benchmarks, which is +18.5% over the second-best baseline (OPD).
• Minimal Computational Overhead: The divergence metric used in SOD reuses log-probabilities already computed in the forward pass, making the distillation process efficient.

Ideal Use Cases

• Small Language Model Agents: Excellent for applications requiring agentic reasoning and tool integration where model size is a constraint.
• Complex Problem Solving: Suited for tasks involving multi-step reasoning in domains like mathematics, science, and code generation.
• Resource-Constrained Environments: Provides strong performance in a compact form factor, making it suitable for deployment on devices with limited computational resources.