WizardKM:Empowering Large Language Models to Follow Complex Instructions

Introduction

The author shows that the instruction data in the current nlp community is relatively simple, and most of them are summarization and translation tasks. However, in real scenarios, people have various needs, which limits the versatility of the model.
The author mentioned that if the quality of this human-labeled qa data is relatively high, it will release the performance of the model very well, but there are some problems in obtaining the data now:

1. Labeling this kind of data is very time-consuming and expensive.
2. Due to the limited professionalism of the labeler, it is difficult to obtain high-quality data.

Based on the above problems, the author proposes a method that can construct a large amount of high-quality data in a short period of time.

As shown in the figure above, the author calls this method Evol-Instruct. Starting from a 1+1 equal to what, the author expands the data from two directions: the depth direction and the breadth direction. Then send the question to ChatGPT to get the qa data pair, and then filter the data.

In order to verify the effectiveness of this method, the data generated by the above method was fine-tuned by Llama7B, called WizardLM, and compared with Alpaca and Vicuna. The author generated 250k pieces of data through the Evol-Instrcut method through the 175 initialization data of Alpaca. For the sake of fairness, the author sampled 70k pieces of data in these data for comparison.

Experiment find-outs:

1. The data generated by Evol-instruct outperforms that of ShareGPT.
2. Under complex test instructions, annotators prefer the output of WizardLM to that of ChatGPT.

reference

https://arxiv.org/pdf/2304.12244.pdf