Multi task prompt learning: how to train a large language model?

title ：Multitask Prompted Learning: How large language models are trained?（ Multi task prompt learning : How to train the big language model ？）

author ：Jing Cheng

 

background

In modern times NLP field , Everything is related to transfer learning . The current model based on neural network has the advantage of scalability , This means that we can simply train larger models on larger data sets . thankfully , We have a mature self supervised learning framework , And the text data on the Internet is very rich —— for example ,Common Crawl The project extracts about... From the web page every month 20TB Text data .

therefore , In recent years ,NLP Researchers focus on developing transfer learning methods . It all started with this paper ： Explore the limitations of transfer learning with a unified text to Text Converter —— link .

In short , It is a unified framework , Form every text processing problem into “ Text to text ” problem ： Give a text sequence as input , The model outputs a text sequence . It allows the same model 、 The goal is 、 The training process and decoding process are applied to each common NLP Mission （ Including translation 、 Question and answer 、 Classification, etc. ）.

 

Prompt

T5 The development of text to text pre training model makes prompting a particularly useful method for multi task learning . As shown above , We first get data sets from different tasks , Then add some guiding descriptions before the original training set . The preprocessed training data is called prompt .

The important thing is that the input is human readable （！）, We've come a long way . Current large NLP The model reads like a human and answers like a human . Think of large pre training models as babies Yoda： It is very powerful for the general knowledge contained in the model W.R.T, But we don't know what kind of behavior （ Tips ） It can trigger superpowers in downstream tasks .

For each data set （ Or a class of tasks ）, It selects a prompt template to transfer the training data to the prompt .

 

Zero sample text generalization

Based on previous work , The paper ：Multitask Prompted Training Enables Zero-Shot Task Generalization, It focuses on explicit training of language models in a supervised and large-scale multitasking manner . It tries to answer these two questions ：

1. Whether multi task prompt training can improve the generalization ability of invisible tasks ？
2. Whether the training of wider prompts will improve the robustness of prompt wording ？

It experiments in this way ： Group data sets by task , Train some groups , And test the model on the reserved data set . By ensuring that no data leakage occurs , It can test the generalization ability of zero samples .

 

What is? “ Mission ”？

We use the term “ Mission ” To refer to the general... Tested by a specific set of data NLP Ability . To evaluate zero sample generalization for new tasks , We train on a subset of tasks , And in one ​​ Evaluate on the task reserved by the Group .

 

 

Model training details

key word ： Encoder - Decoder architecture / Autoregression / Maximum likelihood training

All models we train are based on T5, It's based on Transformer The encoder - Decoder language model , Use from C4 Of 1T Mark up the mask language modeling style target for pre training （Raffel wait forsomeone ,2020）. because T5 The pre training goal of involves filling tags from deleted input text , Therefore, it is completely different from the conditional text generation format used in our prompt dataset . therefore , We use from Lester And others are publicly available LM-adapted T5 Model . （2021）（ be called T5+LM）, It is achieved through the use of models from C4 Of 100B Extra marker training T5 Produced . Unless otherwise stated , Otherwise, we use with 11B Parametric XXL edition .

 

Reference resources

1. Use a unified text to text converter to explore the limitations of transfer learning — https://arxiv.org/abs/1910.10683
2. Multi task prompt training realizes zero sample task generalization — https://arxiv.org/abs/2110.08207
3. Parameter efficient prompt adjusts the scale power — https://arxiv.org/abs/2104.08691

 

source ：

https://medium.com/@chengjing/multitask-prompted-learning-62b87a9b8665