"Smaller Language Models Are Better Instruction Evolvers"

8B parameter models outsmart 70B models in creating training instructions

Small language models (SLMs) can generate more effective instruction tuning data than larger models, challenging the common belief that bigger is better.

-----

https://arxiv.org/abs/2412.11231

🤔 Original Problem:

→ Current approaches rely heavily on large models like GPT-4 (70B+ parameters) for generating instruction data, assuming bigger models are inherently better at this task.

-----

🔍 Solution in this Paper:

→ The researchers compared 8B parameter models (SLMs) against 70B+ models (LLMs) across three instruction evolution scenarios: Evol-Instruct, AutoIF, and Auto Evol-Instruct.

→ They introduced a new metric called Instruction Complex-Aware IFD (IC-IFD) that factors in instruction complexity when evaluating effectiveness.

→ SLMs generate more diverse instructions because they have a broader output space due to lower instruction-following capabilities.

-----

💡 Key Insights:

→ SLMs synthesize more complex and varied instructions than LLMs

→ Second-round SLM instructions outperform third-round LLM instructions

→ SLMs achieve better results with lower computational costs

-----

📊 Results:

→ SLMs consistently outperformed LLMs in instruction following, math reasoning, and code generation tasks

→ SLM-generated instructions showed 6.9% more evolutionary trajectories than LLM-generated ones

→ IC-IFD metric demonstrated more accurate assessment of instruction effectiveness without requiring tuning