Home > Technology peripherals > AI > ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

王林
Release: 2023-05-29 20:29:22
forward
1430 people have browsed it

Recently, the HugAILab team of East China Normal University developed the HugNLP framework, which is a comprehensive and unified NLP training framework for researchers and developers. It can support text classification, text matching, question and answer, information extraction, text Build and train models for various NLP tasks such as generation and small sample learning.

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

##Open source address: https://github.com/HugAILab/HugNLP

Paper: https://arxiv.org/abs/2302.14286

It is worth noting that HugNLP also integrates a large number of the latest Prompt technologies, such as Prompt-Tuning, In-Context Learning, Instruction-tuning, and Chain-of-thought will be introduced in the future

The HugAILab team has also developed a series of applications, such as CLUE&GLUE ranking tools, It can support ChatGPT model training and deployment product HugChat, as well as unified information extraction product HugIE, etc.

HugNLP is a layered framework that follows the development model of "high cohesion and low coupling". Its core includes model layer (Models), processor layer (Processors), and evaluator. There are four parts: Evaluators and Applications.

The framework diagram is as follows:

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

    ## Model layer: Contains the model part, mainly Divide according to task type;
  • Processor layer: load, cache, word segmentation, etc. data, and convert it into Tensor for model input;
  • Evaluator layer: Specify different evaluation processes and evaluation indicators according to different types of tasks (classification or generation);
  • Application layer: Corresponding application execution scripts. Theoretically, selecting a model, a data processor and an evaluator can correspond to an application.
HugNLP is completely developed based on HuggingFace and has easy expansion and deployment capabilities. It also integrates the MLFlow training tracker to facilitate users to track the experimental progress in time and conduct experimental analysis.

The HugNLP framework is called comprehensive because it integrates a large number of NLP task models. The ones that have been implemented include:

    Pre-training: Masked LM, Causal LM, knowledge-enhanced pre-training;
  • Instruction-Tuning: supports unified paradigm training such as autoregressive generative, interval extraction, NLI;
  • Text classification/matching: traditional Fine-tuning, Prompt-tuning, In-Context Learning;
  • Sequence annotation: Supports NER and other sequence annotation tasks ;
  • ##Meta-learning: sequence-based meta-learning (SentenceProto), interval-based meta-learning (SpanProto), token-based meta-learning (TokenProto, NNShot);
  • Q&A: Supports extractive Q&A, multiple-choice Q&A, and open generative Q&A;
  • Text generation: supports text summary, machine translation (under development );
  • Code intelligence: Currently integrated with code tasks such as code clone detection (Clone) and code defect detection (Defact);
Quickly deploy the HugNLP framework

, you only need to execute three lines of code:

git clone https://github.com/HugAILab/HugNLP.gitcd HugNLPpython3 setup.py install
Copy after login
The following introduces several core capabilities of HugNLP:

Benchmark one-click ranking;
  • Pre-training and knowledge injection;
  • Fine-tuning & Prompt-tuning;
  • Instruction-tuning;
  • ##In-Context Learning;
  • half Supervise Self-training;
  • Code code intelligence;
  • 1. Benchmark one-click ranking
HugNLP It was the first to develop ranking tools for some common rankings, such as GLUE, CLUE, etc. Users only need to configure the corresponding data set name to achieve one-click refresh.

为了验证框架的有效性,在22年9月提交了CLUE榜单的刷榜结果,选择一系列中文小模型(RoBERTa、MacBERT、P-BERT等)并结合了logits集成方法,至今依然维持在第15名位置,曾一度超越了部分企业。

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

例如如果训练CLUE榜单的AFQMC数据集,可编辑文件

applications/benchmark/clue/clue_finetune_dev.sh
Copy after login

修改参数:

--user_defined="data_name=afqmc"
Copy after login

执行下列命令即可:

bash applications/benchmark/clue/clue_finetune_dev.sh
Copy after login

同样的方法还可以训练一些常用的NLP任务,例如阅读理解、实体识别、以及GLUE英文数据集等。

HugNLP还集成了一系列模型用于刷榜,例如BERT、RoBERTa、DeBERTa、MacBERT、Erlangshen等。

二、预训练与知识注入

传统的一些预训练模型(例如BERT、GPT2等)是在通用语料上训练的,而对领域事实知识可能不敏感,因此需要显式的在预训练阶段注入事实知识。

HugNLP实现了多个知识增强预训练技术,其中包括DKPLM和KP-PLM。可分解的知识注入方法DKPLM和将结构化知识转化为自然语言形式的注入方法KP-PLM是两种不同的注入方式。由于这些知识注入方法采用的是可插拔式的设计,因此无需改变模型结构,这使得在下游任务上进行微调非常容易。

执行下面命令即可进行Masked Language Modeling和Causal Language Modeling的预训练:

bash applications/pretraining/run_pretrain_mlm.shbash applications/pretraining/run_pretrain_casual_lm.sh
Copy after login

三、 Fine-tuning & Prompt-Tuning

Pre-training和Fine-tuning模式通常被遵循,以基于预训练语言模型的NLP。HugNLP也包含Fine-tuning技术。

3.1 参数有效性学习

HugNLP集成了包括Prefix-tuning、Adapter、BitFit、LoRA等参数有效性训练方法,可以加速模型的训练,降低显存占用量。

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

在训练脚本中,只需要添加一行参数,即可开启参数有效性训练:

--use_freezing
Copy after login

对于参数有效性方法,HugNLP实现了若干类别的分类模型,如下所示:

CLASSIFICATION_MODEL_CLASSES = { "head_prefix_cls": { "bert": BertPrefixForSequenceClassification, "roberta": RobertaPrefixForSequenceClassification, }, "head_ptuning_cls": { "bert": BertPtuningForSequenceClassification, "roberta": RobertaPtuningForSequenceClassification, }, "head_adapter_cls": { "bert": BertAdapterForSequenceClassification, "roberta": RobertaAdapterForSequenceClassification, }, "masked_prompt_cls": { "bert": PromptBertForSequenceClassification, "roberta": PromptRobertaForSequenceClassification, },  "masked_prompt_prefix_cls": { "bert": PromptBertPrefixForSequenceClassification, "roberta": PromptRobertaPrefixForSequenceClassification, }, "masked_prompt_ptuning_cls": { "bert": PromptBertPtuningForSequenceClassification, "roberta": PromptRobertaPtuningForSequenceClassification, }, "masked_prompt_adapter_cls": { "bert": PromptBertAdapterForSequenceClassification, "roberta": PromptRobertaAdapterForSequenceClassification, }, }
Copy after login

只需要指定下面参数即可,例如选择adapter进行分类:

--task_type=head_adapter_cls
Copy after login

3.2 对抗训练:引入对Embedding的扰动,提高模型的鲁棒性

HugNLP框架集成了若干种对抗训练的方法,其中最简单的对抗方法为FGM算法:

  • 首先计算输入样本(通常为word embedding)的损失函数以及在处的梯度:;
  • 计算在输入样本的扰动量:,其中为超参数,默认取1.0;
  • 得到对抗样本:;
  • 根据得到的对抗样本,再次喂入模型中,计算损失,并累积梯度;
  • 恢复原始的word embedding,接着下一个batch。

在训练时,只需要添加一行参数,即可默认调用FGM算法:

--do_adv
Copy after login

3.3 Prompt-tuning:通过模板来复用预训练目标

传统的Fine-tuning在低资源场景下容易出现过拟合问题,因此复用预训练的目标可以拉近Pre-training和Fine-tuning之间的语义差异。

HugNLP集成了PET、P-tuning、Prefix-tuning等Prompt-Tuning算法,并无缝嵌入在NLP分类任务的模型里。

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

在训练时,只需要指定下面两个参数,即可以开启Prompt-tuning模式,例如选择p-tuning算法:

--task_type=masked_prompt_ptuning_cls--use_prompt_for_cls
Copy after login

四、Instruction-tuning

在构建通用人工智能之前,必须将不同类型的自然语言处理任务进行范式统一,尤其是在大模型时代。HugNLP为此定义了三种统一范式的思想:

  • 万物皆可生成:将所有NLP任务建模为单向自回归生成,例如GPT-3、ChatGPT等;
  • 万物皆可抽取:将所有NLP任务建模为抽取式机器阅读理解;
  • 万物皆可推断:将所有NLP任务建模为自然语言推断;

基于三种不同的范式统一,HugNLP推出两个核心产品,分别是:

  • HugChat:基于生成式Instruction的中小型ChatGPT类模型;
  • HugIE:基于抽取式Instruction的统一信息抽取框架;

4.1 HugChat:基于Causal Language Modeling的生成式对话模型

最近ChatGPT火爆全球,为了让研究者可以训练自己的ChatGPT,HugNLP框架集成了基于生成式Instruction的训练产品——HugChat,其支持各种类型的单向生成式模型的训练,例如GPT-2、GPT-Neo、OPT、GLM、LLaMA等。

在8张V100 32G的条件下,可训练OPT-13B大模型。HugAILab团队公布了大约200万条英文和300万条中文的对话数据,以用于模型训练。例如训练GPT-2(XL),可直接执行脚本:

bash ./application/instruction_prompting/HugChat/supervised_finetuning/run_causal_instruction_gpt2_xl.sh
Copy after login

使用基于HugNLP训练的GPT-2(1.3B)模型可以轻松地完成对话任务。只需要执行如下命令即可玩转HugChat:

python3 applications/instruction_prompting/HugChat/hugchat.py
Copy after login

例如可以写套磁信邮件:

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

再例如搜索谷歌地球的相关信息:

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

也可以实现编写简单的代码(1.3B的模型具备此能力已经很惊叹了!):

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

HugNLP目前正在开发其他类型的Decoder-only大模型,相关信息和开源内容如下表所示:

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

HugChat后期将推出垂直领域的大模型解决方案,同时将与OpenAI API进行融合,推出大模型服务框架。

4.2 HugIE:基于Global Pointer的统一信息抽取框架

信息抽取(Information Extraction)旨在从非结构化的文本中抽取出结构化信息,是构建知识库的重要步骤之一。通常信息抽取包括两个核心步骤,分别是命名实体识别(Named Entity Recognition)和关系抽取(Relation Extraction)。

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

我们基于HugNLP研发一款HugIE产品,旨在实现统一信息处理。其主要核心包括如下几个部分:

  • 将实体识别和关系抽取,统一为新的范式——基于抽取式阅读理解的方法。HugIE采用Global Pointer模型实现信息抽取;
  • 定义Instruction Prompt,指导模型生成需要抽取的内容;
  • 采用多任务训练的方法训练;

HugIE目前已经开源了模型:https://huggingface.co/wjn1996/wjn1996-hugnlp-hugie-large-zh 可以基于HugNLP框架使用HugIE抽取模型,如下图所示:

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

五、In-Context Learning

In-Context Learning(ICL) 首次由GPT-3提出,其旨在挑选少量的标注样本作为提示(Prompt),从而在形式上促使大模型生成目标答案。ICL的优势在于无需对参数进行更新,即可实现惊艳的效果。

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

HugNLP框架集成了ICL,主要涉及到样本的挑选和预测结果的校准两个部分:

  • 样本挑选:默认为从训练集中随机挑选样本,后期将会开发一系列样本挑选的算法,例如聚类、K近邻、余弦相似度等;
  • 预测校准:由于所挑选标注样本与待预测样本存在分布差异,需要对预测的概率分布进行校准,这里采用Calibrate Before Use方法,如下图,可以对预测分布进行校准,提高预测效果。

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

目前ICL已经集成在HugNLP里,只需要指定下面参数即可:

--user_defined="data_name=xxx num_incontext_example=4 l=1 use_calibrate=True"--use_prompt_for_cls
Copy after login

六、半监督Self-training

半监督旨在同时结合标注数据和无标签数据来训练NLP任务。Self-training是一种简单但有效的迭代式训练方法,其通过Teacher模型先获取伪标签,对伪标签进行去噪后,再训练Student模型。Self-training方法传统上存在着较多噪声,可能会削弱训练结果。

为了提高性能,HugNLP引入成熟的Uncertainty-aware Self-training技术。框架图如下所示:

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

其采用了来自贝叶斯推断中的MC Dropout技术,即对Teacher模型执行 次推理,每次推理开启Dropout开关,从而得到若干与Teacher模型满足独立同分布的模型预测。

基于这些预测结果,可以通过信息熵的变化量得到Teacher模型对无标签数据的不确定性量化指标(即BALD算法),核心公式如下:

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

进行多次DC Dropout的代码实现如下(详见hugnlp_trainer.py):

y_T = list()for i in tqdm(range(T)): y_pred = [] for step, inputs in enumerate(unlabeled_dataloader): _, logits, __ = self.prediction_step(model, inputs, prediction_loss_only, ignore_keys=ignore_keys) y_pred.extend(logits.detach().cpu().numpy().tolist()) predict_proba = torch.softmax(torch.Tensor(y_pred).to(logits.device), -1) y_T.append(predict_proba.detach().cpu().numpy().tolist()) y_T = np.array(y_T)#compute mean y_mean = np.mean(y_T, axis=0)BALD算法实现如下:def get_BALD_acquisition(y_T):expected_entropy = - np.mean(np.sum(y_T * np.log(y_T + 1e-10), axis=-1), axis=0)expected_p = np.mean(y_T, axis=0)entropy_expected_p = - np.sum(expected_p * np.log(expected_p + 1e-10), axis=-1)return (entropy_expected_p - expected_entropy)
Copy after login

HugNLP使用半监督模式,只需要做两件事:

(1)执行脚本时添加参数:

--use_semi
Copy after login

(2)在指定的数据集目录下,存放unlabeled data文件。

七、其他更丰富的应用

HugNLP has developed numerous applications as listed below, and there are many more exciting applications currently under development.。HugNLP欢迎有志之士加入HugAILab参与开源开发工作。

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training

ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training


The above is the detailed content of ChatGPT models can be directly trained! East China Normal University and NUS open source HugNLP framework: refresh the rankings with one click and fully unify NLP training. For more information, please follow other related articles on the PHP Chinese website!

Related labels:
source:51cto.com
Statement of this Website
The content of this article is voluntarily contributed by netizens, and the copyright belongs to the original author. This site does not assume corresponding legal responsibility. If you find any content suspected of plagiarism or infringement, please contact admin@php.cn
Popular Tutorials
More>
Latest Downloads
More>
Web Effects
Website Source Code
Website Materials
Front End Template