Source: WizardLM / Evol-Instruct (Xu et al., 2023)
Microsoft paper that uses an LLM to rewrite seed instructions into progressively harder ones (Evol-Instruct) to build complex instruction data.
Contents
- What: research paper (arXiv:2304.12244, ICLR 2024)
- Author: Can Xu, Qingfeng Sun, Kai Zheng, Xiubo Geng, Pu Zhao, Jiazhan Feng, Chongyang Tao, Qingwei Lin, Daxin Jiang
- Date: 2023-04-24 (v3 2025-05-27)
- Original: arXiv:2304.12244
Summary
Human-written instruction data skews easy. Evol-Instruct starts from seed instructions and has an LLM rewrite them step by step into more complex ones. The evolved data, mixed together, is used to fine-tune LLaMA into WizardLM.
Key claims
- An LLM, not human annotators, raises instruction complexity through iterative rewriting.
- Human evaluators rated Evol-Instruct instructions above human-created ones.
- On the high-complexity part of the test set, evaluators preferred WizardLM’s outputs to ChatGPT’s.
- In GPT-4 automatic evaluation, WizardLM reached over 90% of ChatGPT’s capability on 17 of 29 skills, while still trailing in some areas.