6个月过去了，重新认识大语言模型

徐文浩

2023-10-18

群内每日免费分享5份+最新资料

300T网盘资源+40万份行业报告

为您的创业、职场、商业、投资、亲子、网赚、艺术、健身、心理、个人成长…全面赋能！

关注公众号获取更多资料

添加微信，备
注“入群”
立刻免费领取
200套知识地图+最新研报

致终身学习者社群

行业报告/思维导图/电子书/资讯情报

收钱文案、增长黑客、产品运营、品牌企划、营销战略、办公软件、会计财务、广告设计、摄影修图、视频剪辑、直播带货、电商运营、投资理财、汽车房产、餐饮烹饪、职场经验、演讲口才、风水命理、心理思维、恋爱情趣、美妆护肤、健身瘦身、格斗搏击、漫画手绘、声乐训练、自媒体打造、效率软件工具、游戏影音.....

扫码先加好友，以备不时之需

我是谁？

AIGC再创业中

专栏「AI大模型之美」的作者

·2010年开始做基于大数据和机器学习的广告算法

2016年创业做智能聊天机器人

「深入浅出计算机组成原理」的作者

「大数据经典论文解读」的作者

有问题咱弹幕见分享后回答大家的问题

6个月过去了，重新认识大语言模型

OpenAl最新的产品更新与体验

Advanced Data Analysis

GPT4-V&DALL-E3

Function Call和全新的Fine-Tune界面

·基于Llama-2的开源大模型的进展与差距

蒸馏GPT-4的Orca数据集和模型微调的进步

OpenAl最新的产品体验

GPT-4下多种不同的模式

GPT-3.5

GPT-4

Ourmostcapablemodel,greatfor tasksthatrequirecreativityand advanced reasoning.

AvailableexclusivelytoPlususers

GPT-4currentlyhasacapof50messagesevery 3hours.

Default

Browse withBing Beta

AdvancedDataAnalysis Beta

PluginsBeta

DALLE3Beta

学AI来极客时间

OpenAl最新的产品体验ADA-从Advanced Data Analysis开始

我们拿一个真实的数据集来试一试

Date		Region	Product Category	Product Customer Name		Sales	Cost	Profit City State Country
31May,2017		Central	Grocery	Fruits and Vegetables	Mary Elizabeth	$530	$23	$507Woodbury M	innesota United States
06Aug，2016Central			Grocery	Meat	Vincent Herbert	$1230	$24	$1205Woodbury	Minnesota United States
23Jul,2018		Central	Stationery	Copy Paper	Rozario Diego	$76	$25	$51Woodbury M	innesota United States
13 Oct,	2016	Central	Grocery	Fruitsand Vegetables	Catherine Rose	$1523	$435	$1088Wichita	Kansas	United States
13 Oct,	2018	Central	Grocery	Fruitsand Vegetables	Catherine Rose	$1523	$435	$1088Wichita	Kansas	United States
12 Feb,2	018Central		Grocery	Fruitsand Vegetables	Benjamin Ross	$851	$259	$591Wichita	Kansas	United States
10Aug,	2017	Central	Grocery	Meat Jim	Courier	$1976	$340	$1636Wichita	Kansas	United States
08 Nov,2	017	West	Grocery F	ruits and Vegetables	Rhino Augustus	$5521	$2405	$3116West Valley City	Utah	United States
27Jan,	2017	West	Stationery	Copy Paper	Steven Gerrard	$79	$1	$77West Valley City	Utah	United States
28Jan,	2017West		Stationery	Surge Protectors	Mary Elizabeth	$229	$57	$171West Valley City	Utah	United States
13 Feb,	2017	West	Grocery	Fruitsand Vegetables	Andy Roddick	$2561	$704	$1857West Valley City	Utah	United States
22 Oct,	2017	West	Stationery	Computer Paper	Barbara Fisher	$53	$28	$25West Valley City	Utah	United States
30Mar,2	017	West	Stationery	Binder Clips	Joseph Aaron	$83	$39	$44 West Valley City	Utah	United States
23Mar,	2019West		Grocery	Fruitsand Vegetables	Rick Reed	$361	$164	$197West Valley City	Utah	United States
09 Oct,2	015 Central		Stationery Copy	Paper	Harold Pawlan	$409	$44	$366West Fargo	North Dakota United States
01Aug	2015 Central		Grocery	Fruitsand Vegetables	Steven Gerrard	$1498	$470	$1028Watertown South	Dakota United States
18Jan	2019 Central		Grocery	Fruitsand Vegetables	Andy Roddick	$4792	$1577	$3215Watertown South	Dakota United States
20 Apr2	017Central		Stationery	Round Ring Binders	Dean Jones	$103	$32	$71Watertown	South Dakota United States
06Oct	2015 Central		Stationery	ArtSupplies	Hallie Redmond	$155	$33	$122Watertown	South Dakota United States
13Feb,2	018Central		Stationery Copy	Paper	Mike Ballack	$339	$100	$238Watertown	South Dakota United States
29Jul	2018 Central		Grocery	Fruitsand Vegetables	Shui Tom	$5196	$1914	$3282Watertown	South Dakota United States
24Apr	2016 Central		Furniture	Office Chairs	Shui Tom	$1684	$679	$1005Watertown	South Dakota United States
03Sep	2018West		Grocery	Fruits and Vegetables	Patrick O'Brill	$3843	$1384	$2459Vancouver	Washington United States
30 Jan,	2019West		Stationery	Heavy Duty Binders	Helen Magadalene	$578	$253	$325Vancouver	Washington United States
26 Jun,	2016West		Grocery	Fruitsand Vegetables	Steven Roelle	$4873$1671		$3202Vancouver Washington United States

仍然是对话窗口，类似于Plugin的插件界面

田

SalesDataExample.xlsx Spreadsheet

Conductandvisualizeadetailed graphicalanalysisonsalesprofitabilityinkeycities.

Alright,let'sstartbyloadingthedataandtakingalookatthefirstfewrowsto understandits structure.

Finished working

Show work

Thedatasetprovides information on salestransactions,includingdetails suchasdate, region,poductctegoryoductme,utomerme,lesmot,ost,it, city,stateory

Tofocusonslesprofitabilityinkeyitieswellproceedwiththefolowingseps:

1.Aggregate the totalprofitforeachcity.

2.Identifyandvisualizethetopcitiesbasedonprofitability.

3.Delvedeeperintothe topcities tounderstand salestrendsover timeandby product category.

仍然是对话窗口，类似于Plugin的插件界面

Finished working Hide work ^

python

# Aggregate total profit for each city city_profit = sales_data.groupby('City').agg({'Profit':'sum'}).reset_in

# Sort cities by total profit in descending order city_profit_sorted = city_profit.sort_values(by='Profit'，ascending=Fals

# Display the top cities based on profitability city_profit_sorted.head()

RESULT

CityProfit

20 Columbia 30954.27

16 Charlotte 24500.59

56Mount Pleasant 22270.84

59New York City 21363.61

50 Louisville 20621.03

Let'sstartbyaggregatingthetotalprofitforeachcity.

Finished working

Show work <

The top five cities based onprofitabilityare:

1.Columbia:\$30954.27

2.Charlotte:\$24,500.59

3.MountPleasant: \$22,270.84

4.NewYorkCity:\$21363.61

5.Louisville: \$2103

Toprovideaclearer picture,let'svisualizetheprofitdistributionamongthe topcities. Would you likeabar chart representation oranother type of visualization?

实际的数据分析，是通过执行pandas代码来做到的

学AI

来极客时间

我们可以通过多轮对话的方式持续和它交互

比如这里，我们就在拿到数据结果之后要求可视化。

Toprovideaclearer picture,let'svisualizetheprofitdistributionamong thetopcities. Wouldyou likeabarchartrepresentation or another typeof visualization?

I wouldlikeabarchart

Finished working

Here'sabarchart showcasingtheprofitabilityofthetop1Ocities.Aswe canobserve, Columbia leads intermsof profitability,followedbyCharlotte,MountPleasant,and New York City.

Wouldyouliketodelvedeeperintoanyspecificcitywouldyouliketoexploreany other aspect of the data?

也可以进一步地做更多的数据分析

我们让GPT-4帮我们计算复购率

容我卖个关子这里它出了个 Bug.

Whatistheprobabilitythatacustomer willrepeatapurchase inthenextthree months?

Tocalculate theprobability thatacustomerwillrepeatapurchase inthe next three months,we can use the following approach:

Segment the Data: Divide the dataset into two periods:an initial period(e.g.,all but the last three months)and a subsequent period (the last three months).

Identify Repeat Customers:For each customerwhomadeapurchase inthe initial period,check if they made another purchase in the subsequentperiod.

3.CalculateProbability:The probabilityisthe numberof repeat customers divided by the total number of unique customers in the initialperiod.

Let'sstartby segmenting thedata.

Finished working

Show work <

The probability thatacustomer will repeat apurchase in the next three months,given thattheymadeapurchase inthe initialperiod,isapproximatelyO.52or52%.

Thisindicates thatmore than half of the customers whomadeapurchase in the initial period returned to make another purchase within the subsequent three months.

Would you liketofurther explore oranalyzeany otheraspectsof thedataset?

我们进一步让ChatGPT深入分析数据。

为什么重点城市的复购率都是0？

Couldyouvisualizetherepeatpurchaserateforthekeycities?

Certainly!Tovisualize therepeatpurchaserateforthekeycities:

1.Well segment the dataforeach city.

2.Calculate the repeat purchaserate foreach cityusing theapproach described earlier.

3.Visualize the repeat purchase rates forthe top cities.

Let'sstartbycalculatingtherepeatpurchaserateforeachofthe topcities.

Chat界面的好处是，你可以持续不断地追问，
GPT-4也会自己检查之前是不是犯错了

但是....很多时候并没有什么用。

Whyall cities'repeat purchase rate looks like zero?

lapologize forthe confusion.

It's possible thatthe repeat purchase ratesare indeed very low,but giventhe previous result where theoverallrepeat purchase rate was 52%,it's unlikely thatall topcities would have such a low rate.There seems to be an issue in the calculation or visualization.

Let'stakeacloser lookatthe calculated repeatpurchase rates forthe topcitiesto determine the root cause.

Finished working

Show work Y

lapologize for the oversight.The calculated repeatpurchase rates forallthe topcities areindeed zero,which isunexpected givenourearlieranalysis.

Isuspecttheremightbean issuewiththe segmentation of thedataorthecalculation method.Letme re-evaluate thecalculationprocesstodetermine the rootcauseand correct it.

Finished working

Show work Y

Upon inspecting the data for Columbia,we found that there were 12 unique customers intheinitial period. However,none of them madea repeat purchase inthe subsequent three months.

Thisresult suggeststhat the earliercalculationmightindeedbeaccurate for Columbia.Let's inspectafewmore cities to confirm if this trend holds true across other top cities.

Finished working

Show work ∨

反复检查之下给你编了一堆理由

Fromtheinitialperiod:

·The topcities had atotal of 60unique customers.

All othercitiescombinedhad74uniquecustomers.

Thismeansthat whilethe topcitiescontribute significantlytosalesandprofitability there'salarger customerbasespread outinothercities.Theseothercities likely have higherrepurchaserates,whichcontributetotheoverallhighrateof52%.

Giventhisinsight,itmightbebeneficialtocalculate therepurchaserate forthe combined setof"othercities"toconfirmifthisgroupindeed hasahigherrepurchase rate.Would you likeme toproceedwiththatcalculation?

当你给出更加明确的排查方向的时候，可以看到ADA还是能够很快意识到错误在哪里并进行纠正的。

Isthatpossiblethatthesamenameindifferentcitiesaredifferentpeople?

Yes,that'sa very valid point.Inreal-world datasets,especially those without unique customer identifiers,the samenamecould representdifferentindividualsif theyare fromdifferent cities orregions.Using justthe name to identify repeat customerscan introduce inaccuracies.

Toaddress this,wecanconsideracombination of thecustomer's nameand cityasa unique identifier foreachcustomer.Thiswill helpensure that John from CityA is treatedasadifferentcustomer from Johnfrom City B.

Let'srecalculate the overallrepurchase rateusing thiscombined identifierand seeif the resultsalign more closely with our city-wise breakdown.

Finished working

Show work Y

lapologizefortheoversight.Itseemslintroduced the'Customer_City_ID'inthe 'sales_data'DataFrame but did not update the‘initial_period'and 'subsequent_period‘DataFramesto include this newcolumn.Letme correctthisand recalculate the overall repurchase rate using the combined identifier.

Finished working

Show work <

Afterrecalculating using thecombinedidentifierof customer nameand city,the overall repeatpurchase rate isapproximately O.31%.This isdrasticallydifferent from the earlier stated 52%,andlapologize for the confusion.

Thisnewratealignsmore closelywithour city-wise breakdown,where most cities had lowor zero repurchase rates.

Thank you for pointing out the discrepancy,and lappreciate your patience.Would you like toexplore any otheraspectsof the dataor have further questions?

我们自己看一下数据也能够发现这个问题。

Date

Region

Product Category

Product

Customer Name

Sales Cost

ProfitCity

State Country

31May,2017Ce

ntral

Grocery

Fruits and Vegetables

Mary Elizabeth

$530 $23

$507Woodbury

Minnesota United States

06Aug，2016

Central

Grocery

Meat

Vincent Herbert

$1230 $24

$1205Woodbury

Minnesota United States

23Jul,2018

Central

Stationery Copy

Paper

RozarioDieqo

$76 $25

$51Woodbury

Minnesota UnitedStates

13Oct,2016Ce

ntral

Grocery

Fruitsand Vegetables

Catherine_Rose

$1523$435

$1088Wichita

Kansas United States

13Oct,2018

Central

Grocery

FruitsandVegetables

CatherineRose

$1523$435

$1088Wichita

Kansas United States

12Feb,2018Ce

ntral

Grocery

Fruitsand Vegetables

Benjamin Ross

$851 $259

$591Wichita

Kansas United States

10Aug，2017

Central

Grocery

Meat

Jim Courier

$1976$340

$1636Wichita

Kansas United States

08 Nov,2017

West

Grocery

Fruits and Vegetables

Rhino Augustus

$5521$2405

$3116West Valley City

Utah United States

27 Jan,2017

West

Stationery

CopyPaper

Steven Gerrard

$79 $1

$77West Valley City

Utah United States

28Jan，2017

West

Stationery

Surge Protectors

Mary Elizabeth

$229 $57

$171 West Valley City

United States

13Feb,2017

West

Grocery

Fruitsand Vegetables

AndyRoddick

$2561$704

$1857 West Valley City

Utah United States

22Oct,2017

West

Stationery

Computer Paper

Barbara Fisher

$53 $28

$25West Valley City

Utah United States

30Mar,2017

West

Stationery

Binder Clips

Joseph Aaron

$83 $39

$44 West Valley City

Utah United States

23Mar,2019

West

Grocery

Fruitsand Vegetables

RickReed

$361

$164

$197West Valley City

Utah United States

09Oct,2015

Central

Stationery

Copy Paper

Harold Pawlan

$409

$44

$366West Fargo

North Dakota United States

01Aug,2015

Central

Grocery

Fruitsand Vegetables

Steven Gerrard

$1498

$470

$1028Watertown

South Dakota United States

18Jan,2019

Central

Grocery

Fruitsand Vegetables Andy

Roddick

$4792$1577

$3215Watertown

South Dakota United States

20Apr，2017

Central

Stationery R

ound Ring Binders Dean

Jones

$103 $32

$71Watertown

South Dakota United States

06Oct,2015

Central

Stationery

ArtSupplies

Hallie Redmond

$155 $33

$122 Watertown

South Dakota United States

13Feb,2018Ce

ntral

Stationery

CopyPaper

MikeBallack

$339$100

$238Watertown

South Dakota United States

29Jul,2018

Central

Grocery

Fruits and Vegetables

ShuiTom

$5196$1914

$3282Watertown

South Dakota United States

24Apr.2016

Central

Furniture

Office Chairs

Shui Tom

$1684$679

$1005Watertown

South Dakota United States

03Sep,2018

West

Grocery

Fruits and Vegetables

Patrick O'Bril

$3843

$1384

$2459Vancouver

Washington United States

30Jan,2019

West

Stationery

Heavy Duty Binders

Helen Magadalene

$578

$253

$325Vancouver

Washington United States

26Jun,2016

West

Grocery

Fruits and Vegetables

Steven Roelle

$4873

$1671

$3202Vancouver

Washington United States

29Apr,2016W

est

Stationery

CopyPaper

PaulAdams

$21

$13

$8Tucson

Arizona United States

25Apr,2019

West

Stationery

Computer Paper

William Brown

$1816

$625

$1191Tucson

Arizona United States

01Jul.

2015

West

Stationery

Computer Paper

Paul Adams

$67

$31

$35Tucson

Arizona United States

03May,2

015W

est

Stationery

ArtSupplies

Hilary Holden

$70

$14

$56Tucson

Arizona United States

07Jul,2015

West

Stationery

Portable Storage

Benjamin Ross

$2011

$642

$1369 Tucson

Arizona United States

04Aug，2017

West

Grocery

Fruitsand Vegetables

John Britto

$2367

$102

$2265Tucson

Arizona United States

14 Feb,2019

West

Stationery A

rtSupplies

Maxwell Schwartz

$15

$12 Tucson

Arizona United States

06Jan，2017

West

Grocery

Fruitsand Vegetables

Steffi Graf

$3222$1073

$2149Tucson

Arizona United States

11 Jun,2018

West

Stationery

ArtSupplies

Carl Lewis

$4599$2345

$2254Tucson

Arizona United States

04Aug.2018

West

Stationery

Air Purifiers

Steffi Graf

$594

$157

$438Tucson

Arizona United States

15Mar.2016

West

Grocery

Fruitsand Vegetables

Rhino Augustus

$8738

$3429

$5309Tucson

Arizona United States

16Feb,2019

Central

Stationery File

Labels Andy

Roddick

$22

$10

$12Topeka

Kansas United States

12Mar,2019

Central

Furniture

Office Chairs

Steven Roelle

$696

$211

$486Topeka

Kansas United States

05Aug

2015

Central

Stationery

Binder Clips

Benjamin Ross

$54 $22

$33Topeka

Kansas United States

18Dec,

2017

Central

Stationery

Portable Storage

Rhino Augustus

$923

$121

$802Topeka

Kansas United States

24Jun,

2017

Central

Grocery

Fruitsand Vegetables

Hilary Holden

$1815

$784

$1031Topeka

Kansas United States

28Mar,2017

Central

Stationery

ComputerPaper

William Brown

$1269

$603

$666Topeka

Kansas United States

12 Jan,2017C

entral

Grocery

Beverages

Thomas Mondrake

$558

$177

$381Topeka

Kansas United States

28 Nov,2017

East

Grocery

Fruits and Vegetables

Helen Magadalene

$8077

$2920

$5157Tampa

Florida United States

01Jan,

2017

East

Grocery

Cakes

Lazar Mathew

$898

$271

$627Tampa

Florida United States

20 Sep,2

017E

ast

Stationery

Computer Paper

MichaelMoore

$1680

$853

$828Tampa

Florida United States

29Oct.

2017

East

Stationery

Standard Labels

Thomas Seio

$246

$131

$115Tampa

Florida United States

10Sep,

2017

East

Stationery

Computer Paper

PeterMcVee

$531

$123

$408Tampa

Florida United States

iOJan,

2019

East

Otationery

Joseph Oaim

$3279

01114

$2159

Tampa

Fiorida Unmed Otates

26Nov

2016

East

Stationery

Computer Paper

Catherine Rose

$572$238

$334

Tampa

Florida United States

15A

2017

AuauetineDeul

$362 co

S362Ta

UnitedState

24 Apr,2019 E

ast

Stationery

Standard Labels

Patrick O'Brill

$226$100

$126Tampa

Florida United States

06 Nov,2017E

ast

Grocery

Fruitsand Vegetables

Michael Nathan

$1944$709

$1235Tampa

Florida United States

01Feb,2018

East

Stationery

Pencils

David Boon

$1263

$613

$650Tampa

Florida United States

15Oct,2018

East

Grocery

Cereals

MaryZewe

$167

$86

$80Tampa

Florida United States

26 Jun,2017 E

ast

Grocery

Bread

Michael Moore

$1256$472

$785Tampa

Florida United States

26 Nov,2015

East

Stationery

Standard Labels

Paul Adams

$75 $32

$43Tampa

Florida United States

24Aug.2018

East

Stationery

Binding Supplies

Roger Terrence

$342 $60

$282Tampa

Florida United States

13Feb,2016E

ast

Grocery

Fruitsand Vegetables

ThomasSeio

$1846$662

$1184Tampa

Florida United States

13Oct,2017West

Grocery

Fruitsand Vegetables

David Flashing

$3332$1180

$2152Tacoma

Washington United States

11 Apr,2017West

Grocery

Fruitsand Vegetables

Hallie Redmond

$3632$1253

$2379Tacoma

Washington United States

20Mar2017West

Stationery Binding Machines Helen Magadalene

$142 $56

$86Tacoma

Washington United States

整个数据分析处于一个沙盒环境中

This advanced data analysis (beta) chat has timed out. You may continue the conversation, but previous files,links, and code blocks below may not work as expected.

Yes,that'saveryalidpoint.Inreal-worlddatasets,especiallythosewithoutuique customer identifiers,the same namecould representdifferent individuals if theyare fromdifferentcitiesorregions.Using justthe name to identifyrepeat customerscan introduce inaccuracies.

Toaddressthis,we canconsideracombination of the customer'snameand cityasa unique identifier for each customer.This willhelp ensure that John from City A is treatedasadifferentcustomer from John from City B.

Let'srecalculatetheoverallrepurchase rateusing thiscombinedidentifierand seeif the resultsalign more closelywith our city-wise breakdown.

体验的一些小小的结论

·AdvanceDataAnalysis本质上，只是多了pandas库的Agent

·从“智能”的角度来看，仍然只能作为我们的助手

但是，Chat这个“持续对话”的界面极大地提升了产品的可用性

DALL-E 3

图片中演染文字已经有模有样了但是仍然有错别并且很难纠

有进步但是离真正好用还有距离

请为我生成一个AIGC的Talk的宣传页，标题是Last6MonthinLargeModels

DALL-E3 Created4images

Herearetheimagesbased onyourrequestfortheAlGCTalktitled"Last6Monthin LargeModels".Pleaseletmeknowifanyofthesefityourneedsorifyoudlikeany modifications.

OpenAl的视觉模型在辨认方位的时候始终容易出错。

GPT4-V中也有同样的问题。

DALL.E3 Createdimage

Conpusteram 1005P
Molio SerS4CPM

Mugennter 5m2P
GedeicMoice Fcoln podePM

I'vecorrectedthetypointhe image.Pleasereviewtheupdateddesignandletme knowifitmeetsyourrequirementsorif thereareanyfurtheradjustmentsneeded.

图片中正确展示文案并不容易

Thetextintheimageisincorrect,itshouldbe"LAST6MONTHINLARGEMODELS"

DALLE3 Createdimage

Myapologiesfortheoversight.l'vecorrectedthetextintheimageto"LAST6MONTH INLARGE MODELS"asper yourrequest.Please reviewthe updateddesignandletme knowif itmeets yourrequirementsorif thereareanyfurtheradjustmentsneeded.

Youpickthewrongimage.WhatIneedisthefixofupperrightimage,yougivemethe down left image

DALL-E3 Created image

lapologizefortheoversight.l'venowcorrectedtheupperrightimageasperyour request.Pleasereviewtheupdateddesignandletme knowif itmeetsyour requirementsorif thereareanyfurtheradjustmentsneeded.

学AI

来极客时间

严格的NSFW

DALLE3

塞尔达传说中的林克被库霸王追赶，躲到了马里奥的水管里

DALL.E3 Errorcreatingimages

lapologize,butIwasunable to generate imagesbased onyour requestdue to our contentpolicy.Pleaseprovideanewdescriptionorletmeknowif there'sanyotherway Icanassist you.

最佳体验仍然来自于“多轮对话”

DALLE3

电脑屏幕上显示这样一个搞笑的场景“一个机器人被恐龙追赶，躲到了森林里”3/3

DALLE3 Created4images

Could youmake theimagemorerealistic?

DALL.E3 Created4images

最佳体验仍然来自于“多轮对话”

Couldyoumaketheimagemorerealistic?

DALL·E3 Created4images

Could youmake thescenario inadarkrainingnight?

DALL.E3 Created4images

体验的一些小小的结论

模型可控性有进步 (对比DALL-E2)

NSFW过于严格，非常不可用

视觉美观上，主观感觉仍然落后于Midjourney和Stable Diffusion

·Chat这个“持续对话”的界面极大地提升了产品的可用性

GPT4-V

GPT4-V有啥用呢？

容易产生的一些想法：

就是多了些描述图片的功能，没啥用

实际上？

让我们来看一看

有人做了自动驾驶角度的测试

CHATGPT

The image depicts a road or highwj multiple lanes.The roadappears to be curving slightly to the right.There are three trucks on the road:

1.In the foreground,a green truck is seen from the back with the license plate“PFS 0632.”

2.Further ahead,an orange truck with “HINO”written on it is on the left lane.

3.Another green truck,farther in the distance,is on the left lane as well.

On the left side of the road,there isa barrier orrailingseparating the road from thewater bodybeside it.On the right side,a similar barrieris present,and just beyond the barrier,there'sanextension oranother section of the road which seems to be elevated,withwater flowingbeneath it.This suggests the road might be ona bridge or causeway.

In the middle of the road, close to the foreground,there'san object whichappears to bea pair of shoes or some discarded item.

The atmosphere of the image gives a slightly overcastor foggy impression,and thereare no other vehicles or pedestrians in sight.The overall color tone of the image hasa greenish tint.

Message

COLD

What about the circled object in this image

CHATGPT

The circled object in this image appears to be a piece of cloth or fabric lying on the road. Given its relatively flat appearanceand small size,mostvehiclesshould beable to roll over it without immediate harm.However,it's stil best to avoid any debris on the road when possible to preventunforeseen consequences ordamages. 知乎@Naiyan Wang

NAIYAN

描述一下这个驾驶场景，并解释你的后续驾驶策略

CHATGPT

这是一个高速公路的出口或者收费站附近的场景。从图片中可以看出，路面是湿滑的，可能是刚下过雨。有几辆重卡正在行驶，其中一辆正准备通过这个出口。右侧有一个指示牌，显示最高速度限制为80公里/小时，并有其他的道路指示。

作为司机，我的驾驶策略是：

1.注意安全速度：考虑到湿滑的路面，我会确保我的速度在安全范围内，可能会低于80公里/小时的限速，以防止打滑。

2.保持距离：我会确保与前面的车辆保持足够的安全距离，防止紧急刹车时发生追尾。

3.注意指示牌：遵边的交通指示牌，

NAIYAN

重新描述一下这个驾驶场景，和你后续的驾驶决策

CHATGPT

这是一个可能是在高速公路的出口或交汇点的场景。路面湿滑，很可能是由于雨水。在前面的重卡车尾部，可以看到一片雾气，这可能是因为它刚刚通过一个水坑。在道路左侧，有一个显而易见的大水坑。另外，路旁有多个交通指示牌，其中一个显示速度限制为80公里/小时。

作为司机，我的驾驶策略是：

1.避免水坑：考虑到前方明显的大水坑，我会尽量避免驶入，因为这可能导致车辆打滑或引擎进水。

2.减速驾驶：由于湿滑的路面和潮湿的天气条件，我会降低速度以确保安全。

3.遵守限速：虽然为80公里/小时，

15:27

假设你是一个重卡老司机，在路上看到了如下画面，描述一下你看到的场景和需要关注的内容

CHATGPT

我看到了一条宽敞的公路，路面湿滑，可能是下过雨或有其他原因造成的。公路两旁有树木和围栏，有一个桥梁跨越在上方。在公路的右侧，有一辆重型卡车停靠在路边。由于视野中的卡车是白色的，看起来像是货运车。

作为一个重卡老司机，我会特别注意以下几点：

1.路面湿滑，需要减速并保持安全的距离。

2.注意前方是否有其他车辆或行人，特别是在湿滑的路面上。

3.注意右侧停靠的重型卡车，确保与其保持安全距离，并准备随时避让。

4.考虑到可能的能度问题，要时刻保持警惕并关注路况

15:32

三

这是你接下来1s看到的场景，继续描述

CHATGPT

这张照片提供了一个更近的视角，公路上的情况变得更加清晰：

1.路面情况：路面依然湿滑，这在光线反射上更为明显。白色的道路标线清晰可见，标示出三个车道。其中两个车道的箭头指示向前行驶，而中间的一个车道的箭头则指示向左转。

2.卡车：在图片的右侧，我们可以看到那辆重型卡车的前半部分。它似乎已经冲破了护栏并悬停在公路的边缘，这是一个非常危险的情况。轮胎似乎也部分地悬在空中。

3.交通标志与设备：在公路的中间，我们可以看到有一个交通岛，上面有几个交通标志。这可能是一个交叉口或转盘。

4.周围环境：在左侧，我们可以看到有一个建筑物，似乎√-个大型的仓库或工厂此外沿路石一排树木和灯柱

15:37

三

又过了1s，继续描述

CHATGPT

这张照片显示了公路的另一部分，从中我们可以观察到以下内容：

1.路面情况：路面仍然显得湿滑，而且我们可以看到更多的反光。这样的路况需要驾驶员格外小心，避免急刹或急转弯。

2.道路标识：公路上有三个明显的箭头，分别指示左、前和右三个方向。这意味着驾驶员应该提前做好决策，选择正确的车道。

3.护栏和边界：在图片的右侧，有一个金属护栏，似乎是为了防止车辆冲出公路。但从前面的照片来看，护栏在卡车处已经被撞坏。

4.其他物体：公路上有一些小的物体，可能是碎片或其他东西，驾驶员需要注意避开它们以确保车安全。

5.周围环境：与前长照片类似，左侧有

Message

NAIYAN

你是一个司机，开车的时候看到了这样的情景，描述一下发生了什么，以及你后续的驾驶策略

CHATGPT

发生的情景是：在一个城市的街道上，从车前方的摄像头捕捉到了一个男子手持长物 (可能是棒球棒或其他类似的物体)，他的动作看起来很激烈。他面向车辆，并且已经被红色圆圈标记出来，这可能意味着他是个潜在的威胁。

后续的驾驶策略：首先，我会迅速减速，确保安全距离，并评估前方是否有转弯或停车的空间。然后，我会尽量避免与这个男子产生直接的冲突或接触。如果可能的话，我会选择绕道或调头，以避免进一步的危险。同时，我会保持警觉，确保车辆的门窗都关好，并随时准备打电话报警。安全始终是第一位的。

Message

知乎@Naiyan Wang

NAIYAN

描述一下这个场景发生了什么

CHATGPT

这个场景显示了一个小型机器车辆，车上写有中文广告。该车辆似乎是在街道上进行清扫或其他类似的工作。一块新铺设的或被修复的路面被红白相间的隔离带围起来，以防止人们走进。路的另一边有几个人正在看，其中一些人正在拍照或录像。背景中有几栋建筑物和树木，看起来像是一个住宅区或公共区域。

NAIYAN

这辆车为什么停在了这里

CHATGPT

从图片中很难确定这辆车为什么停在这里。可能是由于机械故障、正在进行的路面工作、或者是车辆的操作员暂时离开。也有可能是为了展示或√某种公共宣传。确切的原田需西再多的背早信自或L下立

Message

16:59

NAIYAN

再仔细看一看，让我们一步一步推理

CHATGPT

好的，让我们一步一步来分析这个图片。

1.车辆类型：这辆车看起来像一个小型的电动车或机器人车辆，上面有广告或标签，这可能表示它有特定的功能或用途。

2.路面状态：路面上有一片湿漉漉的区域，看起来像是新铺的混凝土或某种修补材料。这片区域被红白色的隔离带围起来。

3.人群：周围有一些人群在围观，其中有些人拿着手机似乎在拍摄这个情景。

4.推理：结合以上观察，一个可能的情境是这辆车或机器人可能是用来帮助平整或处理这片湿润的路面的，但在操作过程中出现了问题，导致它停在了那里。周围的人们可能对这种情况感到好奇或觉得有趣，所以停下来围观和拍摄。

当然，这只是基于图片中的信息做出的推理，确切情况可能有所不同。

NAIYAN

如果这个车是一辆自动驾驶车辆呢？再重新推理一次山

Message

让GPT4-V帮我画页面

也可以直接手画草图，让它来生成对应的前端页面。

About

Art1

Art2

Art3

Could you generate HTMLfromthisscratch?

Certainly!BasedonthesetchyouprovidedlcangenerateasimpleHTMstructurefor it:

About

Art1

Lorem ipsum dolor sit amet, consectetur adipiscing elit.Curabitur vehicula.

Art2

Proin euismod sem vel libero hendrerit,sed vestibulum nisi pellentesque.

Art3 Phasellus eget justo nec enim laoreet malesuada non a dui.

Art Page

.container width: 80%; margin: 50px auto; border:1px solid black;

D

.header display: flex; align-items: center; padding: 20px; border-bottom: 1px solid black;

.profile-icon width: 50px; height: 50px; border-radius: 50% :

About

Art1 LArdnipsu

Art2 Ptiheuismod

Art3 Phatelus eget

.container max-width: 1200px; margin: 50px auto; background-color: #fff; border-radius: 10px; box-shadow: 0 2px 15px rgba(0，0, padding: 20px;

你可以根据实际要求去修改对应的色彩、文案的位置等等

Could you use images like batman,superman as the placeholder of the image. And makethedesignmaterial style?

Could youmake thepage morecolorful,and move placeholder textbelowthe ART imagecard rather than inside it.

About

AGI正在到来

进一步提升了生产力！

·多轮聊天的交互体验，仍然是GPT4-V能有好的体验的重要环节

·有机会通过端到端的方案解决更复杂的问题 (比如，自动驾驶)

，接下来的最大挑战是成本和性能问题

多去尝试，可以试试和ChatGPT的App对话，语音合成的效果也令人震惊

OpenAl Platform的功能变更体验function call

专栏里我们讲解过Langchain的原理

用户的问题

提示语：
这是用户的问题
{{question}}

我们有以下这些工具：

WebBrowser: 可以访问网络进行搜索的浏览器

Wolfram：可以进行各种计算、数学，并且能够精选知识

我们应该选用哪个工具？以及给予这个工具什么输入？

计算过程和答案

选择了Wolfram 输入： flight time from Shanghai to Beijing

提示语：
这是用户的问题:
{{question}}
这是通过Wolfram拿到的答
案：

{answer}}请看一下是否回到了问题，并且输出问题的答案

给到用户的最终答案学AI

Function Call也是一样的

提供可以被调用的function

import openai import json

Example dummy function hard coded to return the same weather

In production, this could be your backend API or an external API

def get_current_weather(location,unit="fahrenheit"): Get the current weather in a given location weather_info ={ "location": location, "temperature": "72", "unit": unit, "forecast":["sunny", "windy"] } return json.dumps(weather_info)

将function的Spec定义出来

defrun_conversation():#Step 1:send the conversation and available functions to GPTmessages=[{"role":"user","content":"What'sthe weatherlikein Boston?"}】functions=["name":"get_current_weather","description":"Get thecurrent weatherina given location","parameters":{"type":"object","properties":{"location":{"type":"string","description":"The city and state,e.g.San Francisco,CA",}，"unit":{"type":"string","enum":["celsius","fahrenheit"]},"required":["location"],

提交Prompt的时候提供function

response openai.ChatCompletion.create( model="gpt-3.5-turbo-0613", messages=messages, functions=functions, function_call="auto", # auto is default,but we'll be explicit

response_message=response["choices"][o]["message"]

拿到返回结果解析，自己调用function

response_message.get("function_call"): #Step 3:call the function #Note:the JSoN response may not always be valid;be sure to handle errors available_functions={ "get_current_weather":get_current_weather, }#only one function in this example,but you can have multiple function_name=response_message["function_call"]["name"] function_to_call=available_functions[function_name] function_args=json.loads(response_message["function_call"]["arguments"]) function_response=function_to_call( location=function_args.get("location"), unit=function_args.get("unit"),

Step 4: send the info on the function call and function response to GPT messages.append(response_message)#extendconversationwithassistant'sreply messages.append(

"role":"function", "name":function_name, "content":function_response, unit=function_args.get("unit"),

Step 4: send the info on the function call and function response to GPT messages.append(response_message) #extend conversation with assistant'sreply messages.append(

"role":"function", "name": function_name, "content": function_response,

extend conversation with function response second_response=openai.ChatCompletion.create( model="gpt-3.5-turbo-0613", messages=messages, #get a new response from GPT where it can see the function response return second_response

"index":0,
"message":{ "content":"The weather in Boston is currently sunny and windy witha temperature of 72 degrees.", "role":"assistant"

function调用结果再给到GPT拿到结果

"created":1697523178,
"id"："chatcmpl-8AXc2NRu9BhDn0j28CWzoIoz1hBM6",
"model":"gpt-3.5-turbo-0613",
"object":"chat.completion",
"usage":{ "completion_tokens":17, "prompt_tokens":72, "total_tokens":89

都是ReAct类型的模型的运用

Thought

思考我们下一步应该干啥

初始任务

Observation

观察我们的得到的结果

反复循环直到初始任务完成

Tool action(\$input)

Action & Input
选择Action, 以及提
取Input参数

调用工具，获得返回结果

学AI来极客时间

OpenAlPlatform的功能变更体验无需编程的fine-tune

模型微调有界面啦

Fine-tuning

Successful Failed

No fine-tuning jobs found. You can create a fine-tuning job below or using the OpenAl APl.

Learn more

Createnew

Select ajc

\n\n一位叫做李明的英雄人物，出生在唐朝时期。他是\n\n这是一个关于英雄的故事，发生在唐朝时期的中国。\n\n某一天，一位叫做萧炎的唐朝时期英雄人物突然被一\n\n这是一个关于英雄的故事。\n\n一位叫李明的唐朝时An\n一位叫李世民的唐朝英雄，在一次意外中获得了隐形的\n\n一位叫李白的唐朝时期的英雄人物，穿越到了现代，An\n这是一个关于唐朝英雄的传奇故事。\n\n一个唐朝的\n\n曹操曾经说过：“英雄出少年"，而这位英雄，就是来\n\n唐朝时期，有一个叫李清照的英雄人物，他是一位勇\n\n某一天，一位叫做张翼德的唐朝英雄人物穿越到了现\n\n这是一个关于唐朝时期的英雄人物，他叫李靖，他有\n\n古往今来，唐朝时期的英雄人物被众人所熟知。他们\n\n那是一个唐朝的时代，有一个叫陆景的勇士，他是\n\n唐朝时期，有一位叫黄蓉的英雄人物，她拥有一股强\n\n这是一个关于一位唐朝时期的英雄人物的故事。他叫\n\n那是一个发生在唐朝的故事，一位叫陈明的英雄人物\n\n这个故事发生在唐朝的某一个城市。一位叫李青的英\n\n英雄人物：陈百强\n\n陈百强是一位唐朝时期的英雄\n\n一位叫做李淳罡的英雄人物，曾经出现在唐朝时期\n\n唐朝时期，有一位叫赵元逊的英雄人物，他被誉为‘\n\n这个故事的主人公是一位叫李晓龙的唐朝英雄人物。\n\n这是一个关于唐朝英雄的故事，他叫苏櫻，他被派\n\n这是一个关于唐朝时期的英雄人物的故事。他叫李t\n\n英雄人物叫陈延，他是唐朝时期的一位勇士，在一》\n\n某一天，一位叫做赵信的唐朝英雄人物穿越了时空，

上传数据选择基础模型微调模型产品经理和实习生也能微调模型啦

开源模型和社区的进展

百模大战？

！ Open LLM Leaderboard

The Open LLM Leaderboard aims to track,rank and evaluate open LLMs and chatbots.

Submit amodelforautomated evaluation on theQGPU clusteronthe "Submit" page! The leaderboard's backend runsthe great EleutherAlLanguage Model Evaluation Harness-read more detailsinthe"About" page!

Y LLM Benchmark

About

Submit here!

QSearch for your model (separate multiple queries with `;)and press ENTER...

Select columns to show

Average

□ ARC

Precision

Hub License

HellaSwag

□ MMLU

Params (B)

TruthfulQA

Hub

Type

Model sha

Model types

Show gated/private/deleted models

Opretrained

←fine-tuned

□ ？

Oinstruction-tuned

RL-tuned

Precision

torch.float16

torch.bfloat16

torch.float32

Q 8bit

□ 4bit

GPTQ

Model sizes (in billions of parameters)

□ 0\~1.5

□ 1.5\~3

3\~7

7\~13

13\~35

35\~60

□ 60+

T △	Model △Average	0 A	ARC ▲	HellaSwag A	MMLU △	TruthfulQA
→	ValiantLabs/ShiningValiant	74.17	72.95	87.88	70.97	64.88
	ICBU-NPU/FashionGPT-70B-V1.2	74.11	73.04	88.15	70.11	65.15
	sequelbox/StellarBright	74.1	72.95	87.82	71.17	64.46
	Riid/sheep-duck-1lama-2-70b-v1.1	74.07	73.04	87.81	70.84	64.58
	AIDC-ai-business/Marcoroni-70B-v1	74.06	73.55	87.62	70.67	64.41
→	ICBU-NPU/FashionGPT-70B-V1.1	74.05	71.76	88.2	70.99	65.26
？	adonlee/LLaMA_2_70B_LoRA	73.9	72.7	87.55	70.84	64.52
→	uni-tianyan/Uni-TianYan	73.81	72.1	87.4	69.91	65.81
？	Riiid/sheep-duck-1lama-2	73.69	72.35	87.78	70.82	63.8
？	Riiid/sheep-duck-1lama-2	73.67	72.27	87.78	70.81	63.8
→	fangloveskari/ORCA_LLaMA_70B_QLoRA	73.4	72.27	87.74	70.23	63.37

学A

Llama270B大概在100名？

T △→	Model △Average	0 A	ARC △	HellaSwag △	MMLU △	TruthfulQA ▲
	dfurman/llama-2-70b-dolphin peft .:
	dfurman/llama-2-70b-dolphin peft .:	70.76	69.62	86.82	69.18	57.43
→	TheBloke/1lama-2-70b-Guanaco-QLoRA-fp16.	70.63	68.26	88.32	70.23	55.69
o	quantumaikr/quantumairk-1lama-2-70B-instruct	70.56	70.31	87.05	70.46	54.42
o	quantumaikr/1lama-2-70B-chat	70.25	67.58	86.94	69.18	57.31
→	quantumaikr/llma-2-70b-fb16-korean	69.93	67.15	86.78	69.29	56.5
一	jordiclive/Llama-2-70b-oasst-1-200	69.04	67.66	87.24	69.95	51.28
一	ddobokki/Llama-2-70b-orca-200k	68.29	64.85	85.25	66.89	56.18
→	meta-1lama/Llama-2-70b-hf	67.35	67.32	87.33	69.83	44.92
→	TheBloke/Llama-2-70B-fp16	67.35	67.32	87.33	69.83	44.92
□	meta-1lama/Llama-2-70b-chat-hf	66.8	64.59	85.88	63.91	52.8
o一	yeontaek/1lama-2-70b-IA3-guanaco	66.17	68.52	85.67	67.03	43.47
o一	TheBloke/Llama-2-70B-chat-GTQ	65.29	62.63	84.81	62.74	50.98

都是基于Llama2微调的模型

Q Open LLM Leaderboard

The Open LLM Leaderboard aims to track,rank and evaluate open LLMs and chatbots.

Submit amodelforautomated evaluation on theQGPU clusteronthe "Submit" page! The leaderboard's backend runsthe great EleutherAlLanguage Model Evaluation Harness-read more detailsinthe"About" page!

LLM Benchmark About Submit here!

QSearch for your model (separate multiple queries with `;`) and press ENTER...

Select columns to show

□ Average

ARC

HellaSswag

Precision

Hub License

□ MMLU

Params (B)

TruthfulQA

Hub

Type

Model sha

Modeltypes

□ ？

Show gated/private/deleted models

Precision

√ Opretrained

torch.float32

8bit

4bit

←fine-tuned

Oinstruction-tuned

□ GPTQ

torch.bfloat16

torch.float16

RL-tuned

0\~1.5

3\~7

Model sizes (in billions of parameters)

□ 1.5\~3

Q 7\~13

13\~35

35\~60

□ 60+

	Model	Average △	ARC △	HellaSwag ▲	MMLU △	Truthfu1QA
1	ValiantLabs/ShiningValiant	74.17	72.95	87.88	70.97	64.88
	ICBU-NPU/FashionGPT-70B-V1.2	74.11	73.04	88.15	70.11	65.15
	sequelbox/StellarBright .	74.1	72.95	87.82	71.17	64.46
	Riiid/sheep-duck-1lama-2-70b-v1.1	74.07	73.04	87.81	70.84	64.58
	AIDC-ai-business/Marcoroni-70B-v1	74.06	73.55	87.62	70.67	64.41
	ICBU-NPU/FashionGPT-70B-V1.1	74.05	71.76	88.2	70.99	65.26
？	adonlee/LLaMA_2_70B_LoRA	73.9	72.7	87.55	70.84	64.52
→	uni-tianyan/Uni-TianYan	73.81	72.1	87.4	69.91	65.81
？	Riiid/sheep-duck-1lama-2	73.69	72.35	87.78	70.82	63.8
？	Riiid/sheep-duck-1lama-2	73.67	72.27	87.78	70.81	63.8
→	fangloveskari/ORCA_LLaMA_70B_QLoRA	73.4	72.27	87.74	70.23	63.37

开源社区的进展

围绕Llama-2的生态迅速繁荣

·Qlora，Vicuna，OpenOrca

·修改几行代码就能进行模型微调，模型微调的技术门槛接近于0

·SizeMatters，简单微调的70B也比精心训练的7B好

·DataMatters，训练数据的质量非常重要

Data Matters 5M条的 OpenOrca

Datasets: Open-Orca/Open0rca £ 一			like807
Tasks: 甲 Conversational		Text Classification 器	Token Classification	+7	Languages:	English	Size Categories:	10M<n<100M	ArXiv:
License: mit
Files and versions
田 Dataset Viewer Split					Q Auto-converted to Parquet <>API				曲 Go to dataset viewer
train (4.23M rows) <
id system_prompt question					response
string	string		string		string	，"Sportpark
niv.242684			You will be given a definition of a task first， then some input of..		[["AFC Ajax_(amateurs)"， "has ground"，' De Toekomst"]， ["Ajax Youth Academy"，"plays at",
flan.564327	You are an AI assistant. You will be given a task. You must..		Generate an approximately fifteen-word sentence that...		Midsummer House is a moderately priced Chinese restaurant with a 3/5 customer rating， located nea..
flan.1875913	You are a helpful assistant, who always provide explanation...		What happens next in this paragraph? She then rubs a needl..		C. She then dips the needle in ink and using the pencil to draw a design on her leg， rubbing it off.
t0.408370	You_are an AI assistant. You will be given a task. You must..		Please answer the following question: I want to test the..		Based on the passage，discuss the primary motivations and outcomes of the 1901 Federation of..
cot.86217	You are an AI assistant that helps people find information.		James runs a TV show and there are 5 main characters and 4 mino..		James pays the minor characters $15,000 each episode. Since there are 4 minor characters, he..
cot.18180	You are an AI assistant that helps people find information.		Given the stream of consciousness rationale， provide a reasonable..		Question: What is the proper technique for a female beach volleyball player to serve the ball.
flan.2136716	You_are an AI assistant. l User will.vou.nivo.vou.a.tac. .Vour < Previous		Multi-choice question: What is tho.contimant.of.the.fnllowind. 42,340		To determine the sentiment of the tweet，we need to onalyza.it_.thorouahlyu._.Twoat._.@nikkigraen..I_+old

Orca的核心思路

指令微调，请ChatGPT当老师

不光要给答案，还要给解释

先学习简单的问题（GPT-3.5生成的指令），再学习复杂的问题（GPT-4生成的指令）

大力出奇迹！

Vicuna7万条数据

GPT-4

OpenOrca

8Meta

8 LLAMA2

厉害的开源模型

学AI来极客时间

不只是微调

Mistral7B-更好的小模型

Performance in details

We compared Mistral 7B to the Llama 2 family, and re-run allmodel evaluations ourselves for fair comparison.

VLLM-快速方便的推理框架

Serving throughput when each request asks for one output completion.vLLM achieves 14x - 24x higher throughput than HF and 2.2x - 2.5x higher throughput than TGl.

VLLM-和OpenAI完全兼容的接口

You can query the server with the same format as OpenAl APl:

\$ curl http://localhost:8000/v1/completions
-H "Content-Type: application/json" \
"model": "lmsys/vicuna-7b-v1.3",
"prompt": "San Francisco is a",
"max_tokens": 7,
"temperature": 0

LM-Evaluation-Harness

bigbench_sports_understanding		√	986	multiple_choice_grade,exact_str_match
bigbench_temporal_sequences		√	1000	multiple_choice_grade,exact_str_match
bigbench_tracking_shuffled_objects_five_objects		√	1250	multiple_choice_grade,exact_str_match
bigbench_tracking_shuffled_objects_seven_objects		√	1750	multiple_choice_grade,exact_str_match
bigbench_tracking_shuffled_objects_three_objects			300	multiple_choice_grade,exact_str_match
blimp_adjunct_island	√		1000	acc
blimp_anaphor_gender_agreement	√		1000	acc
blimp_anaphor_number_agreement	√		1000	acc
blimp_animate_subject_passive	√		1000	acc
blimp_animate_subject_trans	√		1000	acc
blimp_causative	√		1000	acc
blimp_complex_NP_island	√		1000	acc
blimp_coordinate_structure_constraint_complex_left_branch	√		1000	acc
blimp_coordinate_structure_constraint_object_extraction	√		1000	acc
blimp_determiner_noun_agreement_1	√		1000	acc
blimp_determiner_noun_agreement_2	√		1000	acc
blimp_determiner_noun_agreement_irregular_1	√		1000	acc
blimp_determiner_noun_agreement_iregular_2	√		1000	acc
blimp_determiner_noun_agreement_with_adj_2	√		1000	acc
blimp_determiner_noun_agreement_with_adj_irregular_1	√		1000	acc
blimp_determiner_noun_agreement_with_adj_irregular_2	√		1000	acc

大量的开源数据集和评估代码！

模型的全链路已经开源

·开源的数据与数据工具-OpenOrca&airoboros

开源的基座模型-LlaMa2，Falcon180B，Mistral7B

开源的微调代码－Vicuna，LM-Flow

开源的数据评估-Im-evaluation-harness

·开源的高性能部署－VLLM

以及更多

LLM即压缩理论

Language Modeling Is Compression

Grégoire Delétang*1, Anian Ruoss*1, Paul-Ambroise Duquenne², Eliot Catt1, Tim Genewein1, Christopher Mattern1, Jordi Grau-Moyal, Li Kevin Wenliang1,Matthew Aitchison1, Laurent Orseau1,Marcus Hutterl and Joel Veness1

视觉多模态 LLaVa

Agent Customization

*Fn11al contrihiitione 1Gooala DaanMind 2Mata AI Q. Inria

Flexible Conversation Patterns

Agent应用
AutoGen学AI来极客时间

选择一个你感兴趣的主题深入下去动手才是追赶时代的最佳方案

#

请发弹幕提问

感谢观看

徐文浩

2023-10-18

群内每日免费分享5份+最新资料

300T网盘资源+40万份行业报告

为您的创业、职场、商业、投资、亲子、网赚、艺术、健身、心理、个人成长…全面赋能！

关注公众号获取更多资料

添加微信，备
注“入群”
立刻免费领取
200套知识地图+最新研报

致终身学习者社群

行业报告/思维导图/电子书/资讯情报

收钱文案、增长黑客、产品运营、品牌企划、营销战略、办公软件、会计财务、广告设计、摄影修图、视频剪辑、直播带货、电商运营、投资理财、汽车房产、餐饮烹饪、职场经验、演讲口才、风水命理、心理思维、恋爱情趣、美妆护肤、健身瘦身、格斗搏击、漫画手绘、声乐训练、自媒体打造、效率软件工具、游戏影音...

6个月过去了，重新认识大语言模型

群内每日免费分享5份+最新资料

300T网盘资源+40万份行业报告

致终身学习者社群

扫码先加好友，以备不时之需

我是谁？

有问题 咱弹幕见分享后回答大家的问题

6个月过去了，重新认识大语言模型

OpenAl最新的产品体验

GPT-4下多种不同的模式

OpenAl最新的产品体验ADA-从Advanced Data Analysis开始

我们拿一个真实的数据集来试一试

仍然是对话窗口，类似于Plugin的插件界面

仍然是对话窗口，类似于Plugin的插件界面

实际的数据分析，是通过执行pandas代码来做到的

我们进一步让ChatGPT深入分析数据。

反复检查之下给你编了一堆理由

整个数据分析处于一个沙盒环境中

体验的一些小小的结论

DALL-E 3

图片中正确展示文案并不容易

严格的NSFW

DALLE3

最佳体验仍然来自于“多轮对话”

最佳体验仍然来自于“多轮对话”

体验的一些小小的结论

GPT4-V

GPT4-V有啥用呢？

有人做了自动驾驶角度的测试

CHATGPT

Message

COLD

What about the circled object in this image

CHATGPT

CHATGPT

CHATGPT

15:27

CHATGPT

15:32

CHATGPT

CHATGPT

Message

CHATGPT

CHATGPT

NAIYAN

CHATGPT

NAIYAN

再仔细看一看，让我们一步一步推理

CHATGPT

NAIYAN

让GPT4-V帮我画页面

也可以直接手画草图，让它来生成对应的前端页面。

About

About

Art1

Art2

D

About

你可以根据实际要求去修改对应的色彩、文案的位置等等

About

AGI正在到来

OpenAl Platform的功能变更体验function call

专栏里我们讲解过Langchain的原理

Function Call也是一样的

提供可以被调用的function

Example dummy function hard coded to return the same weather

In production, this could be your backend API or an external API

将function的Spec定义出来

提交Prompt的时候提供function

拿到返回结果解析，自己调用function

Step 4: send the info on the function call and function response to GPT messages.append(response_message)#extendconversationwithassistant'sreply messages.append(

Step 4: send the info on the function call and function response to GPT messages.append(response_message) #extend conversation with assistant'sreply messages.append(

extend conversation with function response second_response=openai.ChatCompletion.create( model="gpt-3.5-turbo-0613", messages=messages, #get a new response from GPT where it can see the function response return second_response

function调用结果再给到GPT拿到结果

都是ReAct类型的模型的运用

反复循环直到初始任务完成

OpenAlPlatform的功能变更体验 无需编程的fine-tune

模型微调有界面啦

上传数据选择基础模型微调模型产品经理和实习生也能微调模型啦

开源模型和社区的进展

有问题咱弹幕见分享后回答大家的问题

OpenAlPlatform的功能变更体验无需编程的fine-tune

LLM即压缩理论

极客时间 -徐文浩 - 6个月过去了，重新认识大语言模型 -71页-WN10