> Wang, Yizhong, and Jhu Cs. Instruction Tuning of Large Language Models. # Instruction Tuning of Large Language Models - NLP before 2018: task-specific models - Multi-task learning (MTL) could not generalize to unseen tasks - ==**Instructions enable models to generalize over multiple tasks**== - Significant improvement of LLM performance - Need for expert-written instructions for each task - Most important factors: diverse tasks (rather than more data for single task), bigger models, good instructions or in-context examples - Crowdsourcing is no good because writing instructions requires creativity and expertise - BUT ==**we can use LLMs to generate instructions and instances**==! - GPT-3 gives fairly good data (all fields are valid in 54% of cases) - Self-Instruct boosts GPT-3 by 33.1%, almost as good as InstructGPT - More instructions -> better performance - **OpenAI licensing issues -> use open-source models like LLaMa** ## 🔍 See also - [Super-NaturalInstructions: Generalization via Declarative Instructions on 1600+ NLP Tasks](https://arxiv.org/abs/2204.07705) - [Self-Instruct: Aligning Language Models with Self-Generated Instructions](https://arxiv.org/abs/2212.10560) - GitHub repo : <https://github.com/yizhongw/self-instruct>