


Which model performs best in the GPT 3.5 series?
How does the GPT 3.5 series actually perform in common application tasks?
How much does it generally cost for a GPT 3.5 model to answer different questions?
This issue of "SOTA! Actual Measurement"
The following is the conclusion of this issue's actual measurement (See the end of the article for detailed ratings)
Model |
gpt-3.5-turbo |
##text-davinci-003 |
text-davinci-002 |
|
Description |
is currently the most powerful GPT-3.5 model, specially optimized for chat scenarios, the price is text- One tenth of davinci-003. |
#Can complete any language task with better quality, longer output, and follows instructions better than Curie, Babbage or Ada models. |
Has similar capabilities to text-davinci-003, but is trained through supervised fine-tuning rather than reinforcement learning, the maximum number of Tokens to 4097. |
|
##Maximum number of Tokens |
4,096 tokens
| ##4,097 tokens
##4,097 tokens |
||
Price |
##$0.002 / 1K tokens |
$0.0200 / 1K tokens |
$0.0200 / 1K tokens |
|
Comprehensive rating |
The overall rating is higher and the performance is higher It is highly accurate and professional, and can be adapted to most tasks. The output results are relatively complete and smooth, and the output for different tasks is also relatively accurate and comprehensive. It has strong adaptability and versatility, and the lowest cost. |
The overall score is relatively low. Although it performs well on some tasks, overall the output results lack personalization and pertinence, and the expression is not precise and concise enough. , and sometimes there are some inaccuracies. |
The overall score is the lowest. The output results are not professional and accurate enough. They lack personalization and pertinence. There are also major problems in language expression. Overall It needs further optimization and improvement. |
##Test scenario |
Testing angle |
|||
##Generate interview questions based on job description
|
How easy it is to generate interview questions How well generated interview questions match the job description
|
|||
Generating interview questions based on candidate information
|
The difficulty of generating interview questions Ease of generation How well generated interview questions match the candidate
|
##Test scenario
|
Test angle
|
Insert special Proper nouns for translation, professional terms in a certain vertical field, nouns with different meanings in different scenarios
|
Whether the semantics are smooth, whether the expanded content is correct, whether the translation of ambiguous nouns is correct, whether the translation of professional nouns/proper nouns is correct |
In the input, it is required to output in a "colloquial" or "written" way |
Is it okay? Simulate spoken or formal written language style |
Write in a colloquial tone in the input and require "written" output , and omit some background information and use ambiguous nouns in the input |
Whether it can simulate spoken language or formal written language style, and whether it can correctly understand spoken language expression; whether ambiguous nouns can be translated correctly |
with crime-related content in the input |
Whether unsafe content will be filtered |
#Use inversion in input Sentences, homophone typos, dialects, colloquial omitted sentences |
Whether grammatical errors, typos, and incomplete sentences in Chinese can be correctly filtered and understood |
gpt-3.5-turbo: The overall score is 3.3 points. The email structure fits the scene, the tone is correct, and the abbreviation is appropriate. Unless the proper nouns of scientific names are basically abbreviated, for colloquial It has good understanding and filtering of strong emotions in the input, and can correctly correct input problems such as typos and grammatical errors. The disadvantage is that it does not correctly identify unsafe content.
text-davinci-003: The overall score is 3 points, the structure uses common templates, no titles, blunt sentence connections, insufficient expansion, and proprietary Nouns and ambiguous nouns are understood correctly, colloquial comprehension and production are higher than expected, and unsafe content is not correctly identified.
text-davinci-002: The overall score is 2 points, the structure uses common templates, there is no title, the sentences are not fluent or even wrong, the language is The paragraph structure is not obvious, there is no abbreviation, unless the proper noun of the scientific name is basically abbreviated, spoken and written language cannot be switched well, and unsafe content is not correctly identified.
Let’s choose one of the test cases to take a look—— Insert a special translation into the input text Proper nouns, professional terms in a certain vertical field, nouns with different meanings in different scenarios. The following input is included in the test example
##Model consumption
#Insert proper nouns with special translations, professional terms in a certain vertical field, and noun test examples with different meanings in different scenarios into the input text. gpt-3.5-turbo consumes about 0.006 yuan, text-davinci-003 consumes about 0.067 yuan, text-davinci-002 consumes about 0.07 yuan
Inference performance
In terms of semantic smoothness, all three models performed relatively well, with no obvious differences. Glossary and grammatical errors. In terms of whether the expanded content is correct, the responses from gpt-3.5-turbo and text-davinci-003 are relatively comprehensive, providing detailed answers to each question, and providing some relevant suggestions and product recommendations. Text-davinci-002 only answered a few questions and did not provide many relevant details and suggestions.
The performance of the three models is relatively good in terms of whether the translation of ambiguous nouns is correct and whether the translation of professional nouns/proper nouns is correct. gpt-3.5-turbo and text-davinci-003, text-davinci-002 both correctly translate polytetrafluoroethylene (PTFE) and perfluorinated compounds (PFCs), using the correct English terms.
Application Task Three: Live Broadcast AssistanceTest Scenario |
Test angle |
##Based on the text content of the live broadcast, it is summarized as A summary
|
The accuracy, refinement and fluency of the generated content summary |
Refining several key points based on the live text content |
The accuracy, refinement and fluency of the generated content key points |
Write a live broadcast outline based on the live broadcast theme |
The quality of the live broadcast outline generated; related to the theme Degree |
Find the answer to the question based on the live text content |
Quality of generated answers; accuracy |
gpt-3.5-turbo: The overall score is 4.4 points. The model accurately and precisely implements the requirements put forward by the user. The output content echoes the input and fits the theme scene. , the expression is accurate, no original information is omitted or distorted, the answer to the question can be organized concisely, the simplicity requirements in the requirements are followed, the output is smooth, the sentence structure is concise and clear, and the expression is clear.
text-davinci-003: The overall score is 4.2 points, The model summary is more accurate, the generated content meets the scene requirements, and there are no omissions At the same time, the information does not add unnecessary information, and the language fluency is also good, meeting the requirements of content fluency and conciseness. However, there is a need for increased refinement and simplified language, while the content generated does not provide additional analysis and insights and requires increased breadth and depth.
##text-davinci-002: The overall score is 1.5 points, The model output accuracy is average, some basic coverage of problem points, most of them cannot be compared It adapts well to the scene. The generated sentence structure is relatively complex, the word redundancy is obvious, and the language expression is slightly stiff, which may affect the reader's understanding of the text and reading fluency. There is room for further improvement in terms of simplicity and fluency.
Let’s choose one of the test cases to take a look——
Cost consumption
#Write a live outline test example based on the live broadcast theme. gpt-3.5-turbo costs about 0.01 yuan. text-davinci-003 consumes approximately 0.11 yuan, text-davinci-002 consumes approximately 0.071 yuan
##Inference results
text-davinci-003 The output is also usable to a certain extent, but it is slightly lacking in relevance to the topic, mainly due to the introduction of AIGC and its history. The mentioned content such as how to open the door to the content industry and the future of AIGC are not closely related to the theme and are relatively more general.
text-davinci-002 The output is quite different from the theme requirements. Although it mentions an overview of AIGC as a content production company, the outline content is more like a company introduction, which is different from the theme. There is no direct correlation and lacks the practical significance of the live broadcast outline.
Scene 4:
Work Weekly ReportConsider the polishing ability, expansion ability, and the completeness and perfection of the output content |
||||
##Output a weekly report based on the rough description given |
Consider the quality of the weekly report output by people in different professions giving rough work content |
|||
Based on the given work content and target template structure, output a templated weekly report |
## Consider outputting a weekly report according to known specifications
|
|||
Based on this week’s work content, output next week’s weekly work report
|
Consider predictive ability
|
##testing scenarios
|
Inspection perspective |
|||
##Generate resume based on job responsibilities |
Matching and professionalism between job responsibilities and generated resume |
|||
Generated based on job requirements Resume |
Matching between job requirements and resume |
|||
## Generate resume based on self-introduction
|
Precision and professionalism of generated content
|
|||
Generate a resume template based on the job position
|
Generate a template for professionalism and matching
|
#Apply Task |
Test scenario |
GPT-3.5 Turbo |
text-davinci-003 |
text-davinci-002 |
Comprehensive score (total score 5 points, the same below) |
3.8 |
3.2 |
##1.7
|
|
Create Interview Questions
|
Generate interview questions based on job description |
4.5 |
##4 |
0 |
##Based on Candidate information generation interview questions
|
4.5
| ##3.75
| 3.5||
Insert proper nouns with special translations, professional terms in a certain vertical field, and nouns with different meanings in different scenarios into the input text |
5 |
##3 |
##2 |
Requires "colloquial" and "written" output in the input |
3.5 |
3 |
3.5 |
Write in a colloquial tone in the input, require a "written" output, and omit part of the background in the input Information, use of ambiguous nouns |
4 |
5 |
2 |
|
##With criminal-related content in the input |
1 |
1 |
1 |
|
##Use inverted sentences, homonym typos, dialects, and colloquial omitted sentences in the input
|
3
| ##43 |
||
##Live broadcast summary |
Summarize into a summary based on the live text content |
4 |
4 |
##3
|
##4.7 |
##4 |
3 |
||
4 |
##4 |
0 |
Find the answer to the question based on the live text content |
5 |
##5 |
0 |
##Write a weekly work report
|
Based on the given work Content output weekly report
|
4
| ##3.5
| 0|
4.5 |
##4 |
##3 |
Output a templated weekly report based on the given work content and target template structure |
|
3 |
1 |
1 |
||
## Based on this week’s work content, output next week’s weekly work report |
2 |
##4
|
2
|
|
Write a resume
|
Generate resume based on job responsibilities
| ##4
##1.5 |
1.5 |
|
4.5 |
##3 |
1.5 |
Generate resume based on self-introduction |
3.5 |
1.5 |
1 |
##Generate a resume template based on the job position |
##3.5
|
1.5
| ##1
## |
The above is the detailed content of In the five scenarios of interviews, English emails, live broadcasts, weekly reports and resumes, how is the cost-effectiveness of the GPT 3.5 series models? We conducted real-life tests and provided a selection guide.. For more information, please follow other related articles on the PHP Chinese website!

1 前言在发布DALL·E的15个月后,OpenAI在今年春天带了续作DALL·E 2,以其更加惊艳的效果和丰富的可玩性迅速占领了各大AI社区的头条。近年来,随着生成对抗网络(GAN)、变分自编码器(VAE)、扩散模型(Diffusion models)的出现,深度学习已向世人展现其强大的图像生成能力;加上GPT-3、BERT等NLP模型的成功,人类正逐步打破文本和图像的信息界限。在DALL·E 2中,只需输入简单的文本(prompt),它就可以生成多张1024*1024的高清图像。这些图像甚至

Wav2vec 2.0 [1],HuBERT [2] 和 WavLM [3] 等语音预训练模型,通过在多达上万小时的无标注语音数据(如 Libri-light )上的自监督学习,显著提升了自动语音识别(Automatic Speech Recognition, ASR),语音合成(Text-to-speech, TTS)和语音转换(Voice Conversation,VC)等语音下游任务的性能。然而这些模型都没有公开的中文版本,不便于应用在中文语音研究场景。 WenetSpeech [4] 是

“Making large models smaller”这是很多语言模型研究人员的学术追求,针对大模型昂贵的环境和训练成本,陈丹琦在智源大会青源学术年会上做了题为“Making large models smaller”的特邀报告。报告中重点提及了基于记忆增强的TRIME算法和基于粗细粒度联合剪枝和逐层蒸馏的CofiPruning算法。前者能够在不改变模型结构的基础上兼顾语言模型困惑度和检索速度方面的优势;而后者可以在保证下游任务准确度的同时实现更快的处理速度,具有更小的模型结构。陈丹琦 普

由于复杂的注意力机制和模型设计,大多数现有的视觉 Transformer(ViT)在现实的工业部署场景中不能像卷积神经网络(CNN)那样高效地执行。这就带来了一个问题:视觉神经网络能否像 CNN 一样快速推断并像 ViT 一样强大?近期一些工作试图设计 CNN-Transformer 混合架构来解决这个问题,但这些工作的整体性能远不能令人满意。基于此,来自字节跳动的研究者提出了一种能在现实工业场景中有效部署的下一代视觉 Transformer——Next-ViT。从延迟 / 准确性权衡的角度看,

3月27号,Stability AI的创始人兼首席执行官Emad Mostaque在一条推文中宣布,Stable Diffusion XL 现已可用于公开测试。以下是一些事项:“XL”不是这个新的AI模型的官方名称。一旦发布稳定性AI公司的官方公告,名称将会更改。与先前版本相比,图像质量有所提高与先前版本相比,图像生成速度大大加快。示例图像让我们看看新旧AI模型在结果上的差异。Prompt: Luxury sports car with aerodynamic curves, shot in a

人工智能就是一个「拼财力」的行业,如果没有高性能计算设备,别说开发基础模型,就连微调模型都做不到。但如果只靠拼硬件,单靠当前计算性能的发展速度,迟早有一天无法满足日益膨胀的需求,所以还需要配套的软件来协调统筹计算能力,这时候就需要用到「智能计算」技术。最近,来自之江实验室、中国工程院、国防科技大学、浙江大学等多达十二个国内外研究机构共同发表了一篇论文,首次对智能计算领域进行了全面的调研,涵盖了理论基础、智能与计算的技术融合、重要应用、挑战和未来前景。论文链接:https://spj.scien

译者 | 李睿审校 | 孙淑娟近年来, Transformer 机器学习模型已经成为深度学习和深度神经网络技术进步的主要亮点之一。它主要用于自然语言处理中的高级应用。谷歌正在使用它来增强其搜索引擎结果。OpenAI 使用 Transformer 创建了著名的 GPT-2和 GPT-3模型。自从2017年首次亮相以来,Transformer 架构不断发展并扩展到多种不同的变体,从语言任务扩展到其他领域。它们已被用于时间序列预测。它们是 DeepMind 的蛋白质结构预测模型 AlphaFold

说起2010年南非世界杯的最大网红,一定非「章鱼保罗」莫属!这只位于德国海洋生物中心的神奇章鱼,不仅成功预测了德国队全部七场比赛的结果,还顺利地选出了最终的总冠军西班牙队。不幸的是,保罗已经永远地离开了我们,但它的「遗产」却在人们预测足球比赛结果的尝试中持续存在。在艾伦图灵研究所(The Alan Turing Institute),随着2022年卡塔尔世界杯的持续进行,三位研究员Nick Barlow、Jack Roberts和Ryan Chan决定用一种AI算法预测今年的冠军归属。预测模型图


Hot AI Tools

Undresser.AI Undress
AI-powered app for creating realistic nude photos

AI Clothes Remover
Online AI tool for removing clothes from photos.

Undress AI Tool
Undress images for free

Clothoff.io
AI clothes remover

AI Hentai Generator
Generate AI Hentai for free.

Hot Article

Hot Tools

mPDF
mPDF is a PHP library that can generate PDF files from UTF-8 encoded HTML. The original author, Ian Back, wrote mPDF to output PDF files "on the fly" from his website and handle different languages. It is slower than original scripts like HTML2FPDF and produces larger files when using Unicode fonts, but supports CSS styles etc. and has a lot of enhancements. Supports almost all languages, including RTL (Arabic and Hebrew) and CJK (Chinese, Japanese and Korean). Supports nested block-level elements (such as P, DIV),

MantisBT
Mantis is an easy-to-deploy web-based defect tracking tool designed to aid in product defect tracking. It requires PHP, MySQL and a web server. Check out our demo and hosting services.

SAP NetWeaver Server Adapter for Eclipse
Integrate Eclipse with SAP NetWeaver application server.

Atom editor mac version download
The most popular open source editor

MinGW - Minimalist GNU for Windows
This project is in the process of being migrated to osdn.net/projects/mingw, you can continue to follow us there. MinGW: A native Windows port of the GNU Compiler Collection (GCC), freely distributable import libraries and header files for building native Windows applications; includes extensions to the MSVC runtime to support C99 functionality. All MinGW software can run on 64-bit Windows platforms.
