Sample 8 instructions Collect human-written Use pretrained LLM to determine if| (both human- and LLM-generated) it's a classification task
Instruction : Find out if the given text is in favor of or apainst aboet
Generate possible class label first, then generate input
Given the instruction, let LLM write the input field first, then generate the output Pruning & fitering Add instruction to task pool| if overlap with existing instructions Generate instances (that is, the "outputs" for a given instruction)
3.1 什么是 Backtranslation?
回译在传统的机器学习中是一种数据增强方法,比如从中文翻译成英文,再从英文翻译会中文,这 样生成的中文与原来的中文在语义上是一致的,但是文本不同;然而 SFT 数据生成的回译 (https://arxiv.org/abs/2308.06259)则是通过输出来生成指令,具体步骤如下图所示:
Step 2. Self-Curation. Step 1. Self-Augmentation. Initialization Iteratively select high-quality generate instructions for unlabelled data to create candidate training data iteration self training Instruction-labeling model that can be applied to unlabeled texts Iteration Iteration Augmented Data Unlabelled Data { yi } Llama-v1| LLM
Mo M2
LLaMA Small dataset with Augmented Data Augmented Data instruction- CA output pairs A selected high-quality subset for
Instruction-labeling model trained
on first batch of augmented data
Seed Data the next round
“Humpback" is the model Humpback
WizardLLM
trained with the proposed instruction backtranslation
Alpaca GPT-4
60- method Vicuna (sharegpt) - OA Open Assistant LIMA Alpaca human-generated Against text-davinci-003 10-
Data Size