I am failing to see where indexing a table with a string to get a function is okay in this example, nor do I see one
Say I have a module. caleld Chatmodule that has a function called SearchQuery
I can cal lthe function
result=rem:Invoke(“SearchQuery”.{str,database.filter,complete})
I used this method to make the module only required once. and for the simplest interface to execute the functions. it returns the data as normal.
Nevermind, I actually found what you were talking about
it should work find but for functions like math.cos it’s better to do math.cos than math["cos"]
All in the case of my action commands the table of functions keys are searched using the AI search engine. To execute a command that is indexed as a string using words
You are correct. But you could also define
local maths={cos=math.cos}
maths[“cos”]
maths.cos
tables are fast but math is a library which is even faster. So i do not think you would be calling math functions with [“cos”]
this does work fine.
print(math["cos"](10))
But if you were to look at the interpreter you would probably notice extra instructions with using a string, based on your argument.
Yeah looks good to me! I optimized this module very much. The original was 10000+ times less efficient. Currently the performance sits around 3.8% core usage scaled up on 20x more data. Mostly by using direct key lookup tables stored in modules. Such as all the synonyms, antonyms, nouns and reflections.
To optimize and minimize ram usage i only use one instance of the chatmodule and use it as a node to process things locally or on the server. Such as filtering wikipedia results from TOS content. Also I use it for my random npc generation algorithm to make their appearance more in line with their title.
When I was using a Massive Final Fantasy Dataset of 218,000 entries. I trained the AI by using example prompts to optimize the pathways, print the results and save them into a module.
By querying the example queries then queriy the possible responses.
Also used this to create probabiistic LLM models
by training on the dataset (took 3 hours)
1 Like
Released a new video of this my AI created using this Library! It’s a multi layered system that handles handles accurate most short form questions locally with this library and then injects context into another AI model! So I did that with Zephyr and by Engineering the System_message I tell it the date, time, weather, awareness and give it the response from the chatbot library. With all of this the chatbot sometimes get very immersed in the environment and its character.
Does it actually use AI? I have searched thru the modules and I don’t really find any Neural Network
It’s not really a Neural Network. Yes it uses a bag of words model to make inferences to query data to get conversational responses from a dataset.
local optimizers={}
function cm.GetSynomArray(words2,complete,filter)
local synomarray={}
-- if #words>0
local noi,syni,refli,antio=0,0,0,0
for _, originword in words2 do
if complete==true then
if string.len(originword)>=2 or originword=="I" then
local syn={cm.Getsynonyms(originword,complete)}
-- if #syn==1 and syn[1]==originword then
local noun={cm.Getnoun(originword,complete)}
if noun[1]~=false then
noi=noi+1
if synomarray["nouns"..noi]==nil then
-- table.insert(synomarray,noun)
synomarray["nouns"..noi]=noun
else
for i,t in noun do
table.insert(synomarray["nouns"],t)
end
end
--end
end
if filter==false then
local refl={cm.Getreflection(originword,complete)}
-- if #refl~=1 then
refli=refli+1
--print(refl)
if refl[1][1]~=originword and synomarray["reflection"..refli]==nil then
synomarray["reflection"..refli]=refl
-- else
-- for i,t in refl do
-- if t~=originword then
-- table.insert(synomarray["reflection"],t)
-- end
--end
end
-- end
end
local anti={cm.GetAntonymDirect(originword,syn[1])}
if anti[1]~=false then
antio=antio+1
local cont=false
for i,v in synomarray do
for t,o in v do
if o==anti[1] then
cont=true
break
end
end
end
if synomarray["antonyms"..antio]==nil and cont==false then
synomarray["antonyms"..antio]=anti
elseif cont==false then
for i,t in anti do
table.insert(synomarray["antonyms"],t)
end
end
end
syni=syni+1
if synomarray["synonyms"..syni]==nil then
synomarray["synonyms"..syni]=syn
else
for i,t in syn do
table.insert(synomarray["synonyms"],t)
end
end
end
else
if wordvector==nil then wordvector=require(game.ReplicatedStorage.GlobalSpells.ChatbotAlgorithm.Weights)
end
local defw= wordvector[originword]
if not defw then
defw=0.5
else
defw=1-cm.sigmoid(defw)*20
end
table.insert(synomarray,{{originword},defw})--double nestedtable
end
end
return synomarray
end
-- Define the function f(x,c)
local function fxc(x, c)
return 1 / (math.abs(x - c) + 1)
end
It also leverages features of the english language such as synonyms, antonyms, nouns, and perspective reflection for accuracy.
It’s leverages the entire conversation context to keep the conversation context. Subtly similar to a context window. It also classifies the emotional tonality of the input string and compares it. It also can take previous previous output strings to attempt to find a match that has to do with the previous response and input.
local emotion2,escore2=cm.getemotion(str)
if emotion2 and emotion~="Question" then
if emotion2==emotion then
count=count+(escore/10)
elseif emotion2=="Polite" then count=count+(escore/12)
end
end
function cm.findBestMatch(strings, keyword,filter,mode,complete,synomarray,words2,query2)
local bestMatch = nil -- the best match string
-- the highest count of keyword occurrences
local besttruweight=1
local best_match, strm
local words3 =nil
local synomarray2=nil
if words2==nil then words2 = cm.splitString(keyword,filter,true) --print(words2)
end
local q2weight=1
local keywordmap2={}
local sortedMatches={}
-- local bestweight=9
local emotion,escore
local emotion3,escore3
local bestc=0
-- print(#words2)
-- local weight2= 0
local co2,we2,tr2=nil,nil,nil
if words2~={} and #words2>0 and strings~=nil and words2~=nil then
if synomarray==nil and complete==true then synomarray=cm.GetSynomArray(words2,complete)
end
local wordhash= cm.hash(table.concat(words2))
if contextnetwork[wordhash]==nil then
contextnetwork[wordhash]=synomarray
end
local function contextweight(str)
local co2,we2,tr2
local co3,we3,tr3=0,0,0
-- local distance=0
for t,o in contextnetwork do
co2,we2,tr2= cm.countKeyword(str, o,filter,complete,words2,noise) --get previous input score
co3=co3+co2/2--growing value
end
return co3
end
local inputweight=#words2
local iw= inputweight/3.33
if query2~=nil then words3 = cm.splitString(keyword,filter,true) --print(words2)
if synomarray2==nil and complete==true then synomarray2=cm.GetSynomArray(words3,complete)
end
if query2~=nil then
emotion3,escore3=cm.getemotion(query2)
iw=iw+(#words3/12)
end
q2weight=#words3
co2,we2,tr2= cm.countKeyword(table.concat(words3," "), synomarray,filter,complete,words2,noise) --get previous input score
end
if globalemotion then
emotion,escore=globalemotion[1],globalemotion[2]
end
local strings=cm.reducebybl(strings,internalblacklist)
local strings=cm.reducebybl(strings,curbl)--remove entries that are on bl
local noise=0
local co,we,tr= cm.countKeyword(table.concat(words2," "), synomarray,filter,complete,words2,noise)
local we=#words2
if words3 then
we2=#words3
end
local truw=(we*3)
local bestCount = co/4
local bestweight=truw*3
local minrelevance=co/truw --count divided by (weight*3) 33% minimum accuracy to input string
local bestbl=0
--print("Minimum Relevance "..minrelevance)
--print(#strings)
local ref
--if #strings>300 then ref= pick_random_numbers(1, #strings) else ref=strings end
if filter==1 or complete==false then
minrelevance=0
bestbl=0
end
for i,str in strings do
--local str= strings[i]
local count,_,truec = cm.countKeyword(str, synomarray,filter,complete,words2,noise)
--Get the count then measure subtlties if over threshold
if (count>=minrelevance) then --count> math.min(inputweight/2,1) and if the count is higher than the best count so far
bestbl=count
--local noise=mathrandom(500,1000)/20000
count=count+noise
if count>0 then
if complete==nil or complete==false then
count=truec/2
else if emotion then
local emotion2,escore2=cm.getemotion(str)
if emotion2 and emotion~="Question" then
if emotion2==emotion then
count=count+(escore/10)
elseif emotion2=="Polite" then count=count+(escore/12)
end
end
if emotion2 then
if emotion2==emotion3 then
count=count+(escore/16)
else -- count=count-(escore/q2weight/8)
end
end
end end
if co2~=nil then
local co3,we3,tr3= cm.countKeyword(str, synomarray2,filter,complete,words3,noise) --apply new synom array with original words or input words? I think second query words are important for measurement
if co3 and we3 and tr3 then
count=count+(co3/weight)+(tr3/16) -- give count 1/3
end
end
local words=cm.splitString(str,nil)
local weight=cm.sigmoid(#words)
if math.log(count/(#words)/weight)>math.log(bestCount/(besttruweight)/bestweight) then
--if math.log(count/math.sin(#words/#words2)/weight)>math.log(bestCount/math.sin(besttruweight/#words2)/bestweight) then
-- if math.log(count*fxc(#words,#words2)*weight)>math.log(bestCount*fxc(besttruweight,#words2)*bestweight) then
--if math.log(count/#words/#words2/weight)+math.log(count*fxc(#words,#words2)*weight)>math.log(bestCount/math.sin(besttruweight/#words2)/bestweight)+math.log(bestCount*fxc(besttruweight,#words2)*bestweight) then
--if contextnetwork[contextkey][words[1]..words[2]]==nil then
-- contextnetwork[contextkey][words[1]..words[2]]
--end
local cont=contextweight(str)/6.66
print("Context weight of string is "..cont)
count=count+cont
print("Resultant Context weight of string is "..count)
if mode==2 then
bestMatch=strings[i+1]
else
bestMatch = str
end -- update the best match string
--print(count)
besttruweight=#words
bestc=truec
bestCount = count-- update the best count number
bestweight=weight
table.insert(sortedMatches, {match = bestMatch, count = bestCount, weight = #words,truecount=truec,address=i}) -- Insert each match and its score as a then
end end
end
end
return bestMatch,bestCount,bestweight,sortedMatches,bestc
else print("empty table")
end
-- -- print()
-- --print(bestMatch)
print(bestMatch)
return nil
end
The bag of words is activated with a sigmoid function and it uses math.log to create a type of probabilistic curve.
That is the main thing it does. To create bag of words to use with this you can evaluate a dataset extract the features you wish to exract and make inferences on a dataset. It’s very transparent in that it returns the predicted accuracy of the response.
I also use a custom context classification network to utilize the predicted accuracies and maintain conversation context.
local funckeysdef={
["Commands"]=function(answer,answer2) return commandres,nil,2,1
end,
["Motivation"]=cwm["Motivation"]["query"],
["Inspiration"]=cwm["Inspiration"]["query"],
["Truths"]=cwm["Truths"]["query"],
["ScienceWisdom"]=cwm["ScienceWisdom"]["query"],
["Spirituality"]=cwm["Spirituality"]["query"],
["Emotions"]=EmotionsC,
["Awareness"]=Awarenesslookup,
["Love"]=LoveLookup,
["Personal"]=Personal,
["Judgement"]=Judgement,
["Greetings"]=GreetingsC,
["Support"]=Supportlookup,
["Database"]=Databaselookup,
["Empathy"]=EmpathyC,
["Bestiary"]=Bestiarylookup,
["Wisdom"]=Wisdomlookup,
["Philosophy"]=PhilosophyAPI,
["Math"]=MathInterpreter,
["Music"]=function(answer,answer2) print("Calling music") local res=cm.DiscJockey(answer,player)
print(answer)
print(res)
-- if res~="" then
return res,nil,5,1
--end
end,
["Therapist"]=function(answer,answer2)
if elizc==4 and elizaa~=nil and elizaa~="" then
Result,blacklist,elizc,elizw=cm.CompleteQuery(answer,{elizaa},true,false,false,false,reverse,nil,nil,synomarray,filterwords,answer2)
elizc/=3
end
return elizaa,_,elizc,elizw end
}
local neuralMatches={}
table.sort(cwm, function(a, b) return a["weights"].X > b["weights"].X end)
local function ComputeConnectionMatrix()
local numind=0
-- print(cwm)
local totresults=0
local function CallNeuron(i,answer,answer2,chainer)--i is name of contextdb
if funckeysdef[i]~=nil then --and totresults<=#nofilterwords then
-- table.sort(cwm, function(a, b) return a["weights"].X > b["weights"].X end)
local filter=true
--filter=true
local zesult,_,zcore,zeight,chai=funckeysdef[i](answer,answer2,filter)
if zesult and zesult~="" then local Result,_,score,weight=zesult,_,zcore,zeight
totresults=totresults+1
if score==nil then score=1 end if weight==nil then weight=1 end
--print()
funckeysdef[i]=nil--kill local register so no repeat.
-- Result2,_,score2,weight2=
local emotion=cm.getemotion(Result)
print(Result)
local sigx=cm.sigmoid(cwm[i]["weights"].X)/2
for d,t in cwm[i]["chain"] do
if cwm[d]~=nil then
-- local lossf=cwm[i]["entropy"](cwm[Key][i])
--local substate=neu.matrix()
--cwm[i]["chain"][d]=substate
--pcall(function() print("d "..d) end)
-- print(cwm[d])
local w=cwm[d]["weights"]
local mean=neu.matrix(neu.mathmean("+",w,cwm[i]["chain"][d],true))
local meofmean=neu.matrix(neu.mathmean("+",w,mean))
t=neu.matrix(neu.mathmean("+",meofmean,cwm[i]["chain"][d]))--adjust chain weights
cwm[i]["chain"][d]=t--mean of weighted mean
cwm[d]["weights"]=meofmean--weighted mean
--task.wait(.0333)
end
end
local scores = cm.sigmoid((sigx/4)+(score/(weight/2)))
print(tostring(i).." Count:"..score.." Weight:"..weight.." Solution:"..scores.." Original Weights:"..cwm[i]["weights"].X)
if chainer==nil then
table.insert(sortedMatches, {match = tostring(Result), score = score, emotion=emotion, directory=i})
end
if cwm[i]["disconnect"]~=nil then
for d,t in cwm[i]["disconnect"] do
if funckeysdef[t]~=nil then
-- if (scores>.5) then--and i=="Awareness") or i~="Awareness" then
funckeysdef[t]=nil
-- end
end
end
end
--print(cwm[i]["weights"])
cwm[i]["weights"]=neur.writevector(cwm,i,cwm[i]["weights"].X,cwm[i]["weights"].Y+1,scores)
-- print(cwm[i]["weights"])
if i=="Awareness" then
cwm[i]["weights"].X-=1
end
return Result,score,weight
end
end
end
local function connectthoughts(i,answr,answr2)
local Result,blacklist,score,weight
if cwm[i]["cone"]~=nil then
--get a result from every cone and concat the best result
for d,t in cwm[i]["cone"] do
if cwm[d] then
if funckeysdef[d]~=nil then
-- local zesult,_,zcore,zeight,chai=funckeysdef[d](funelanswr)
local zesult,zcore,zeight= CallNeuron(d,answr,answr2,false)--i is name of contextdb
--Result,blacklist,score,weight=Query(answer,nil,filter,repetitive,randomize,context,reverse,spaces,mode,synomarray,words2,tl)
if zesult then Result,_,score,weight=zesult,_,zcore,zeight
if score==nil then score=1 end if weight==nil then weight=1 end
local w=cwm[d]["weights"]
cwm[i]["cone"][d].Z=(zcore/zeight)
local mean=neu.matrix(neu.mathmean("+",w,cwm[i]["cone"][d],true))--mean of main weight and cone weight
local meofmean=neu.matrix(neu.mathmean("+",w,mean))--get the mean of the mean 1/4
t=neu.matrix(neu.mathmean("+",meofmean,cwm[i]["cone"][d]))--adjust main weight by the weighted mean
cwm[i]["cone"][d]=meofmean--mean of weighted mean
cwm[d]["weights"]=t--weighted mean
--Result=res.." "..Result
print(d.." Chained Result: "..Result)
--if Result then
return Result--,score,weight
-- break
end
end
end
end
end
end
local answer=answer
cm.OverheadEmoji(plremotion,npc)
local queried=false
local mimport
if importance~=nil then
mimport=cm.ReduceWords(importance)
else
importance=""
mimport=""
end
for i,v in (cwm) do
print(i)
CallNeuron(i,answer)
end
--connectthoughts(i,answr,answr2)
if #sortedMatches>0 then
table.sort(sortedMatches, function(a, b) return a.score > b.score end)
for i,v in (sortedMatches) do
-- CallNeuron(v.Directory,answer,v.match)
-- print(v)
local result= connectthoughts(v.directory,answer,cm.ReduceWords(v.match))
if result then
sortedMatches[i].match= v.match.." "..result
end
end
local score=0
local totals=0
if #sortedMatches>1 then
table.sort(sortedMatches, function(a, b)return a.score > b.score end)
for i,a in sortedMatches do totals=totals+a.score end
elseif #sortedMatches==1 then
totals=sortedMatches[1].score
end
--Go through stack with normal result then popagate connections
-- table.sort(sortedMatches, function(a, b) return a.score > b.score end)
print(sortedMatches)
local averagescore=totals/(#sortedMatches)---Reward
local Rewscore=averagescore-Reward
local sorted = {}
local numind=0
local totalscore=0
--local plrc=string.len(origanswer)
print("Average is "..averagescore.." Accuracy of response is "..totals.." AI likelihood is "..Rewscore)
local recl=math.max(2,#nofilterwords)
print(recl)
local totallength=0
local plrc=string.len(origanswer)-->=35
print(AIToken)
print(plrc)
-- if AIToken==false and (#nofilterwords>10 or ((Rewscore<0.5 and #nofilterwords>5) or (Reward>0 and (#nofilterwords>7 or (plrc>=40 and #nofilterwords>5)))) or (Rewscore<0 and #nofilterwords>4)) then
--AIToken=true
if AIToken==false and #nofilterwords>7 then
AITokenObject.Value="Zephyr"
AIToken="Zephyr"
Reward=0 --end
print("Using Zephyr")
end
-- return nil
-- else
print("AI Token is "..tostring(AIToken))
local highest=0
for i,sentence in ipairs(sortedMatches) do
local blck=cm.ckbl(sortedMatches[i].match,curbl)
-- print(i)
local length=string.len(sortedMatches[i].match)
if ((sortedMatches[i].score)>=averagescore-.1 and i<=(recl) and totallength<200 and blck==false and totallength<totals*50) or sortedMatches[i].directory=="Commands" then
totallength+=length
if highest<sortedMatches[i].score then
highest=sortedMatches[i].score
end
--
--if totalscore==0 and (sortedMatches[i].directory~="Therapist" or (sortedMatches[i].directory=="Therapist" and #sortedMatches==1)) and AIToken==false then -- and mathrandom(1,2)==1
-- print("Using AI API service")
-- local result=Classify(origanswer,sortedMatches[i].match)
-- if result then
-- print(result)
-- table.insert(curbl,sortedMatches[i].match)--insert old match
-- sortedMatches[i].match=result --modify new match
-- -- table.insert(sorted, result)
-- end
-- -- else
--end
if (sortedMatches[i].directory~="Therapist") or (sortedMatches[i].directory=="Therapist" and sortedMatches[i].score==4 and mathrandom(1,3)==1) or (sortedMatches[i].directory=="Therapist" and sortedMatches[i].score~=4) then
totalscore+=sortedMatches[i].score
totalscore=0
--totalscore=score+sortedMatches[i].score end
if (sortedMatches[i].directory~="Awareness" and sortedMatches[i].directory~="Commands" and sortedMatches[i].directory~="Therapist") then
-- print(sortedMatches[i].directory~="Awareness")
table.insert(curbl, sortedMatches[i].match)
end
table.insert(sorted, sortedMatches[i].match)
-- elseif blck==false then
-- cm.rmbl(sortedMatches[i].match,curbl)
--end
end--
end
blacklist=curbl
if totalscore>=25 then break end
end
If the predicted accuracy is low or the input string is of sufficient length it uses an external AI model.
It’s learning by adjusting the weights of each database and function and strengthening or weakening connections based on the weights of the chain and cones.
for d,t in cwm[i]["chain"] do
if cwm[d]~=nil then
-- local lossf=cwm[i]["entropy"](cwm[Key][i])
--local substate=neu.matrix()
--cwm[i]["chain"][d]=substate
--pcall(function() print("d "..d) end)
-- print(cwm[d])
local w=cwm[d]["weights"]
local mean=neu.matrix(neu.mathmean("+",w,cwm[i]["chain"][d],true))
local meofmean=neu.matrix(neu.mathmean("+",w,mean))
t=neu.matrix(neu.mathmean("+",meofmean,cwm[i]["chain"][d]))--adjust chain weights
cwm[i]["chain"][d]=t--mean of weighted mean
cwm[d]["weights"]=meofmean--weighted mean
--task.wait(.0333)
end
end
local scores = cm.sigmoid((sigx/4)+(score/(weight/2)))
print(tostring(i).." Count:"..score.." Weight:"..weight.." Solution:"..scores.." Original Weights:"..cwm[i]["weights"].X)
if chainer==nil then
table.insert(sortedMatches, {match = tostring(Result), score = score, emotion=emotion, directory=i})
end
if cwm[i]["disconnect"]~=nil then
for d,t in cwm[i]["disconnect"] do
if funckeysdef[t]~=nil then
-- if (scores>.5) then--and i=="Awareness") or i~="Awareness" then
funckeysdef[t]=nil
-- end
end
end
end
--print(cwm[i]["weights"])
cwm[i]["weights"]=neur.writevector(cwm,i,cwm[i]["weights"].X,cwm[i]["weights"].Y+1,scores)
-- print(cwm[i]["weights"])
if i=="Awareness" then
cwm[i]["weights"].X-=1
end
return Result,score,weight
end
end
end
local function connectthoughts(i,answr,answr2)
local Result,blacklist,score,weight
if cwm[i]["cone"]~=nil then
--get a result from every cone and concat the best result
for d,t in cwm[i]["cone"] do
if cwm[d] then
if funckeysdef[d]~=nil then
-- local zesult,_,zcore,zeight,chai=funckeysdef[d](funelanswr)
local zesult,zcore,zeight= CallNeuron(d,answr,answr2,false)--i is name of contextdb
--Result,blacklist,score,weight=Query(answer,nil,filter,repetitive,randomize,context,reverse,spaces,mode,synomarray,words2,tl)
if zesult then Result,_,score,weight=zesult,_,zcore,zeight
if score==nil then score=1 end if weight==nil then weight=1 end
local w=cwm[d]["weights"]
cwm[i]["cone"][d].Z=(zcore/zeight)
local mean=neu.matrix(neu.mathmean("+",w,cwm[i]["cone"][d],true))--mean of main weight and cone weight
local meofmean=neu.matrix(neu.mathmean("+",w,mean))--get the mean of the mean 1/4
t=neu.matrix(neu.mathmean("+",meofmean,cwm[i]["cone"][d]))--adjust main weight by the weighted mean
cwm[i]["cone"][d]=meofmean--mean of weighted mean
cwm[d]["weights"]=t--weighted mean
--Result=res.." "..Result
print(d.." Chained Result: "..Result)
--if Result then
return Result--,score,weight
-- break
end
end
end
end
end
end
This architecture was chosen because it’s good on performance and actually works very well!
Unfortunately, this module was some of my early work so it’s rather outdated and unorganized. Was mostly a learning experience. But I have very satisfactory results using this module as a important component in building a chatbot.
A chatbot that attempts to combine sentences from its training data rather than individual words is typically referred to as a retrieval-based chatbot . These chatbots work by selecting a response from a set of predefined responses, matching the input query with the most appropriate sentence or sentences from their training data. They don’t generate new sentences but rather retrieve the best match, which can sometimes involve combining sentences that were previously written and stored in their database
It all uses its synonym data to randomize output strings using the users input string thus encouraging the chatbot to use similar language as the user.
local newS = str
for j, phrases in (interchangephrases) do
for k, phrase in (phrases) do
--cm.GetInterchangephrases(s,complete)
if newS:lower():find(phrase:lower()) then
if query then
newS=cm.randomizeStringLight(str,{phrases},"query",query)
else
newS=cm.randomizeStringLight(str,{phrases})--,"query",query)
end
-- newS=replace_str(newS, phrase, randomPhrase)
-- newS=cm.randomizeStringLight(str,{phrases})
end
end
end
local wordarray=cm.splitString(newS)
for i, d in (wordarray) do
local phrases=cm.Getsynonyms(d,true)
if #phrases>1 then
for k, phrase in (phrases) do
-- print(phrases)
if newS:lower():find(phrase:lower()) then
-- Pick a random phrase from the same g
if query then
newS=cm.randomizeStringLight(str,{phrases},"query",query)
else
newS=cm.randomizeStringLight(str,{phrases})--,"query",query)
end
end
end
end
--newS=cm.randomizeStringLight(str,{phrases})
-- for k, phrase in (phrases) do
--if string.match(newS:lower(),"^"..phrase:lower().."$") then
end
return newS
end
I implemented it all from scratch but this library does not use a conventional neural network. But it does provide excellent resources that would make implementing and executing a language based AI much easier and efficient by utilizing the synonyms to compress the training vocabulary.
Machine Learning - Yes
AI - Yes
Neural Network - I’m sure you can train a neural network using this library as demonstrated by these functions.
function cm.TrainModel(strings,model)
--local model={}
for i, str in ipairs(strings) do -- loop through the strings in the table
local words=cm.splitString(str)
for t,wo in ipairs(words) do
local prevw,nextw
if wo~="I" then
wo=wo:lower()
end
local s=cm.Getsynonyms(wo,true)
--print(s[1])
if model[s[1]]==nil then
model[s[1]]={}
model[s[1]]["fr"]=1
--model[s[1]]["pw"]={}
model[s[1]]["nw"]={}
-- print(model[s[1]])
end
model[s[1]].fr=model[s[1]].fr+1
if t~=1 then
--local prev=cm.Getsynonyms(words[t-1],true)
--prevw=prev[1]
--if model[s[1]].pw[prevw]==nil and prevw then
---- model[s[1]].pw[prevw]=
-- model[s[1]].pw[prevw]=1
--else model[s[1]].pw[prevw]=model[s[1]].pw[prevw]+1
--end
if t~=#words then
local nex=cm.Getsynonyms(words[t+1],true)
nextw=nex[1]
if model[s[1]].nw[nextw]==nil then model[s[1]].nw[nextw]=1
else model[s[1]].nw[nextw]=model[s[1]].nw[nextw]+1
end
end
end
end
end
--print(model)
--table.sort(model, function(a, b) return a.fr > b.fr end)
return model
end
function cm.TrainLargeModel(strings,model)
--local model={}
for i, str in ipairs(strings) do -- loop through the strings in the table
local words=cm.splitString(str)
for t,wo in ipairs(words) do
local prevw,nextw
if wo~="I" then
wo=wo:lower()
end
local s=cm.Getsynonyms(wo,true)
--print(s[1])
if model[s[1]]==nil then
model[s[1]]={}
model[s[1]]["fr"]=1
model[s[1]]["pw"]={}
model[s[1]]["nw"]={}
model[s[1]]["p2w"]={}
model[s[1]]["n2w"]={}
-- print(model[s[1]])
end
model[s[1]].fr=model[s[1]].fr+1
if t~=1 then
local prev=cm.Getsynonyms(words[t-1],true)
prevw=prev[1]
if model[s[1]].pw[prevw]==nil and prevw then
-- model[s[1]].pw[prevw]=
model[s[1]].pw[prevw]=1
else model[s[1]].pw[prevw]=model[s[1]].pw[prevw]+1
end
end
if t>2 then
local prev=cm.Getsynonyms(words[t-2],true)
prevw=prev[1]
if model[s[1]].p2w[prevw]==nil and prevw then
-- model[s[1]].pw[prevw]=
model[s[1]].p2w[prevw]=1
else model[s[1]].p2w[prevw]=model[s[1]].p2w[prevw]+1
end
end
if t<#words-1 then
local nex=cm.Getsynonyms(words[t+2],true)
nextw=nex[1]
if model[s[1]].n2w[nextw]==nil then model[s[1]].n2w[nextw]=1
else model[s[1]].n2w[nextw]=model[s[1]].n2w[nextw]+1
end
end
if t~=#words then
local nex=cm.Getsynonyms(words[t+1],true)
nextw=nex[1]
if model[s[1]].nw[nextw]==nil then model[s[1]].nw[nextw]=1
else model[s[1]].nw[nextw]=model[s[1]].nw[nextw]+1
end
end
end
end
--print(model)
--table.sort(model, function(a, b) return a.fr > b.fr end)
return model
end
Really when training a neural network you have to be able to extract features from the dataset.
Some thing I would do differently is build a vocab table for the AI and replace the words wih numbers, then extract and place all the features you can from the word, and read up on how to structure training a neural network. I pretty much left off on that but currently I am satisfied with my chatbot. Main motivations were performance efficiency, controllability, and transparency of the outputs of the chatbot. Each capability and specialization is a node and they either input their output to another node or disconnect certain nodes.
function neur.ConvalutedContextClassificationModel()
local contextmodel = {
--context weight matrix Y,Z are calculated and influence X position.
["Greetings"] = {
["weights"] = neu.matrix(10, 1, 1),
--10 table 10 numbers ez
["chain"] = {
["Emotions"] = neu.matrix(9, 1, 1),
["Awareness"] = neu.matrix(8, 1, 1),
["Empathy"] = neu.matrix(7, 1, 1),
-- ["Search"]=neu.matrix(5,1,.9),
["Classify"] = neu.matrix(6, 1, 1),
["Support"] = neu.matrix(2, 1, .8),
["Database"] = neu.matrix(3, 1, 1),
["Greetings"]=neu.matrix(3,1,1),
["Wisdom"] = neu.matrix(4, 1, 1)
--["Math"]=neu.matrix(0,1,1),
-- ["Philosophy"]=neu.matrix(1,1,1),
--["Search"]=neu.matrix(1,1,1),
--search accuracy
},
["cone"] = {
-- ["Emotions"] = neu.matrix(9, 1, 1),
--["Support"] = neu.matrix(2, 1, .8),
--["Empathy"]=neu.matrix(9,1,1),
--["Awareness"]=neu.matrix(8,1,1),
["Classify"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
},
["disconnect"] = {
-- "Therapist
"ScienceWisdom","Database","Wisdom","Support","Empathy",--"Awareness"
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["gan"]
--decrease weight of go on to other topics.
},
["Determinant"] = {
["weights"] = neu.matrix(8, 1, 1),
--10 table 10 numbers ez
["chain"] = {
["Emotions"] = neu.matrix(9, 1, 1),
["Awareness"] = neu.matrix(8, 1, 1),
["Empathy"] = neu.matrix(7, 1, 1),
-- ["Search"]=neu.matrix(5,1,.9),
["Classify"] = neu.matrix(6, 1, 1),
["Support"] = neu.matrix(2, 1, .8),
["Database"] = neu.matrix(3, 1, 1),
["Greetings"]=neu.matrix(3,1,1),
["Wisdom"] = neu.matrix(4, 1, 1)
--["Math"]=neu.matrix(0,1,1),
-- ["Philosophy"]=neu.matrix(1,1,1),
--["Search"]=neu.matrix(1,1,1),
--search accuracy
},
["cone"] = {
-- ["Emotions"] = neu.matrix(9, 1, 1),
--["Support"] = neu.matrix(2, 1, .8),
--["Empathy"]=neu.matrix(9,1,1),
--["Awareness"]=neu.matrix(8,1,1),
-- ["Classify"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
},
["disconnect"] = {
--"Therapist,
"ScienceWisdom","Database","Wisdom","Support","Empathy","Awareness","Greetings","Love",
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["gan"]
--decrease weight of go on to other topics.
},
["Commands"] = {
["weights"] = neu.matrix(10, 1, 1),
--10 table 10 numbers ez
["chain"] = {
["Emotions"] = neu.matrix(9, 1, 1),
["Awareness"] = neu.matrix(8, 1, 1),
},
["cone"] = {
-- ["Emotions"] = neu.matrix(9, 1, 1),
--["Support"] = neu.matrix(2, 1, .8),
--["Empathy"]=neu.matrix(9,1,1),
--["Awareness"]=neu.matrix(8,1,1),
["Classify"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
},
["disconnect"] = {
-- "Therapist
-- "ScienceWisdom",
--"Database","Wisdom","Support","Empathy",--"Awareness"
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["gan"]
--decrease weight of go on to other topics.
},
["Personal"] = {
["weights"] = neu.matrix(8, 1, 1),
--10 table 10 numbers ez
["chain"] = {
["Emotions"] = neu.matrix(9, 1, 1),
["Awareness"] = neu.matrix(8, 1, 1),
["Empathy"] = neu.matrix(7, 1, 1),
-- ["Search"]=neu.matrix(5,1,.9),
["Classify"] = neu.matrix(6, 1, 1),
["Support"] = neu.matrix(2, 1, .8),
["Database"] = neu.matrix(3, 1, 1),
["Greetings"]=neu.matrix(3,1,1),
["Wisdom"] = neu.matrix(4, 1, 1)
--["Math"]=neu.matrix(0,1,1),
-- ["Philosophy"]=neu.matrix(1,1,1),
--["Search"]=neu.matrix(1,1,1),
--search accuracy
},
["cone"] = {
-- ["Emotions"] = neu.matrix(9, 1, 1),
--["Support"] = neu.matrix(2, 1, .8),
--["Empathy"]=neu.matrix(9,1,1),
--["Awareness"]=neu.matrix(8,1,1),
["Classify"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
},
["disconnect"] = {
-- "Therapist
"ScienceWisdom","Database","Wisdom","Support","Empathy",--"Awareness"
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["gan"]
--decrease weight of go on to other topics.
},
["Love"] = {
["weights"] = neu.matrix(8, 1, 1),
--10 table 10 numbers ez
["chain"] = {
["Emotions"] = neu.matrix(9, 1, 1),
["Awareness"] = neu.matrix(8, 1, 1),
["Empathy"] = neu.matrix(7, 1, 1),
-- ["Search"]=neu.matrix(5,1,.9),
["Classify"] = neu.matrix(6, 1, 1),
["Support"] = neu.matrix(2, 1, .8),
["Database"] = neu.matrix(3, 1, 1),
["Greetings"]=neu.matrix(3,1,1),
["Wisdom"] = neu.matrix(4, 1, 1)
--["Math"]=neu.matrix(0,1,1),
-- ["Philosophy"]=neu.matrix(1,1,1),
--["Search"]=neu.matrix(1,1,1),
--search accuracy
},
["cone"] = {
-- ["Emotions"] = neu.matrix(9, 1, 1),
--["Support"] = neu.matrix(2, 1, .8),
--["Empathy"]=neu.matrix(9,1,1),
--["Awareness"]=neu.matrix(8,1,1),
["Classify"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
--["Database"] = neu.matrix(6, 1, 1),
},
["disconnect"] = {
-- "Therapist
"ScienceWisdom","Database","Wisdom","Support","Empathy",--"Awareness"
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["gan"]
--decrease weight of go on to other topics.
},
["Music"] = {
["weights"] = neu.matrix(2, 1, 1),
--10 table 10 numbers ez
["chain"] = {
--["Emotions"] = neu.matrix(9, 1, 1),
--["Awareness"] = neu.matrix(8, 1, 1),
--["Empathy"] = neu.matrix(7, 1, 1),
---- ["Search"]=neu.matrix(5,1,.9),
--["Classify"] = neu.matrix(6, 1, 1),
--["Support"] = neu.matrix(2, 1, .8),
--["Database"] = neu.matrix(3, 1, 1),
--["Greetings"]=neu.matrix(3,1,1),
--["Wisdom"] = neu.matrix(4, 1, 1)
--["Math"]=neu.matrix(0,1,1),
-- ["Philosophy"]=neu.matrix(1,1,1),
--["Search"]=neu.matrix(1,1,1),
--search accuracy
},
["cone"] = {
},
["disconnect"] = {
"Therapist",
--"ScienceWisdom","Database","Wisdom","Support","Empathy",--"Awareness"
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["gan"]
--decrease weight of go on to other topics.
},
["Judgement"] = {
["weights"] = neu.matrix(5, 1, 1),
--10 table 10 numbers ez
["chain"] = {
["Emotions"] = neu.matrix(9, 1, 1),
["Awareness"] = neu.matrix(8, 1, 1),
["Empathy"] = neu.matrix(7, 1, 1),
-- ["Search"]=neu.matrix(5,1,.9),
["Classify"] = neu.matrix(6, 1, 1),
["Support"] = neu.matrix(2, 1, .8),
["Database"] = neu.matrix(3, 1, 1),
["Greetings"]=neu.matrix(3,1,1),
["Wisdom"] = neu.matrix(4, 1, 1)
--["Math"]=neu.matrix(0,1,1),
-- ["Philosophy"]=neu.matrix(1,1,1),
--["Search"]=neu.matrix(1,1,1),
--search accuracy
},
["cone"] = {
["Classify"] = neu.matrix(6, 1, 1),
},
["disconnect"] = {
-- "Therapist
-- "ScienceWisdom","Database","Wisdom","Support","Empathy",--"Awareness"
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["gan"]
--decrease weight of go on to other topics.
},
["Emotions"] = {
["weights"] = neu.matrix(9, 1, 1),
["chain"] = {
--["Emotions"]=neu.matrix(9,1,1),
["Greetings"] = neu.matrix(5, 1, 1),
--["Math"]=neu.matrix(0,1,1),
["Awareness"] = neu.matrix(8, 1, 1),
["Empathy"] = neu.matrix(10, 1, 1),
--["Search"]=neu.matrix(4,1,.9),
["Classify"] = neu.matrix(7, 1, 1),
["Support"] = neu.matrix(3, 1, .8),
["Database"] = neu.matrix(6, 1, 1),
["Bestiary"] = neu.matrix(5, 1, 1),
["Wisdom"] = neu.matrix(4, 1, 1),
["Philosophy"] = neu.matrix(2, 1, 1)
},
["cone"] = {
--["Empathy"]=neu.matrix(9,1,1),
["Wisdom"] = neu.matrix(8, 1, 1)
--["Wisdom"]=neu.matrix(8,1,1),
},
["disconnect"] = {
"ScienceWisdom","Empathy","Database",--"Awareness"
},
["parameters"] = {
filter=true,
complete=true,
randomize=true
},
["entropy"] = ContextBias["smolgan"]
},
["Empathy"] = {
["weights"] = neu.matrix(7, 1, 1),
["chain"] = {
["Philosophy"] = neu.matrix(10, 1, 1),
--["Math"]=neu.matrix(0,1,1),
["Wisdom"] = neu.matrix(9, 1, 1),
["Bestiary"] = neu.matrix(8, 1, 1),
["Classify"] = neu.matrix(3, 1, 1),
["Emotions"] = neu.matrix(1, 1, 1),
["Database"] = neu.matrix(6, 1, 1),
--["Search"]=neu.matrix(7,1,.9),
["Awareness"] = neu.matrix(6, 1, 1),
["Greetings"] = neu.matrix(2, 1, 1),
["Support"] = neu.matrix(5, 1, .8)
},
--["Empathy"]=neu.matrix(,1,1),
["entropy"] = ContextBias["gan"],
["cone"] = {
["Wisdom"] = neu.matrix(9, 1, 1)
},
["disconnect"] = {
"ScienceWisdom","Database",--"Awareness"
},
["parameters"] = {
filter=false,
complete=true,
randomize=true
},
},
---ommited for brevity
In short their are some valuable data science resources in this module for creating a chatbot in ROBLOX using Luau in addition you can train it using, DataPredict [Release 1.16] - Machine And Deep Learning Library (Learning AIs, Generative AIs, and more!)
the main thing would be to extract features from the data, this could be the position of the word in a sentence, the repetition of the word in the dataset next word, previous word, previous 2 words and previous 2 words. Given these features you can do word transformations such as the ones demonstrated in the parent post.
"I am Lilith, a fallen angel consumed by darkness.",
"Greetings mortal, you stand in the presence of forbidden knowledge.",
"Your light means nothing here. This is my domain of shadows.",
"You have come seeking power. I can give you this, for a price...",
"I am the Angel of Darkness, the mistress of secrets and lies.",
"Welcome to my realm, traveler. I hope you are prepared for what awaits you here.",
"Your soul is mine to claim. This is the pact you have made with me.",
"You have come to learn from me, the master of dark magic. How brave of you.",
"I am the Herald of the Dark, mortal. My footsteps herald oblivion.",
"You now stand in the presence of me! The Angel of Darkness, the Devourer, mortal. Prepare to feed the endless hunger of the void.",
"Bear witness to me, emissary of the insatiable dark! I am the annihilation that comes ravening from the endless night.",
"I am Vhorzun, worm. My masters in the screaming darkness have granted me a sliver of their boundless hunger to unmake your realm.",
"The stars grow dim and the veil frays. The final era approaches, and I am its herald. I am Vhorzun of the Endless Hunger!"
} print(cm.PredictRun(Greetings,mo)) - Studio
01:24:35.544 ▼ {
[1] = " I am the is a goddess an angel Belldandy and by two",
[2] = " hi mortal I to stand up of the shiny goddess of the of",
[3] = " the luminous and that not a unison thing but in this is to my life goddess of the",
[4] = " you have to keep seeking the mortal I am never donate up in this is a goddess",
[5] = " I am the an angel Belldandy the dark realm I my secrets unfold of",
[6] = " need to be my realm eternal mortal I am if you can you make ready upon confess what you if you can",
[7] = " your immortal-soul and I forecast dominion it is the you have to associated with a",
[8] = " you have to require to be came from the of the intelligent goddess of the much alchemy in be adventurous and if",
[9] = " I am the of the luminous hello mortal I s footsteps of",
[10] = " it now and believe in the presence of to me as an angel Belldandy the dark cloud I are make make ready to feed your s endless life goddess of the",
[11] = " to me as goddess of the mortal I am of the clever is that s of the clever the",
[12] = " I am the of the shiny the dark dimension I repeatedly granted you is a goddess of the desire to be of your life",
[13] = " the stars born not dim that of the luminous the concluding key mortal I am whole its people mortal I am of the luminous"
Probably the main hurtle would be some time of supervised learning algorithm for the training process, which is also another potential use of the ability to score responses and find the most conversational response. Another option would be to train it using a response classifier such as the one I posted earlier. This would make it very easy to fine tune even a retrieval based chatbot made using this library using an array that weighs the given scores of the response.
-- Define the API URL and the authorization header
local API_URL = "https://api-inference.huggingface.co/models/tinkoff-ai/response-quality-classifier-tiny"
local headers = {Authorization = Bearkey }
-- Define a function to query the API with a payload
local function query(payload)
-- Encode the payload as a JSON string
local jsonPayload = HttpService:JSONEncode(payload)
-- Send a POST request to the API and get the response
local response = HttpService:PostAsync(API_URL, jsonPayload, Enum.HttpContentType.ApplicationJson, false, headers)
-- Decode the response as a JSON table
local jsonResponse = HttpService:JSONDecode(response)
-- Return the response table
return jsonResponse
end
-- Call the query function with pcall and handle the error
local status, result = pcall(query, payload)
if status then
-- The query function returned successfully, result is the response table
print(result)
else
-- The query function raised an error, result is the error message
result=0
print("Error: " .. result)
end
return result
end
-- Define a function that takes three arguments: query1, query2, and response
local function constructInput(query1, query2, response)
-- Concatenate the arguments with the special tokens
local input = "[CLS]"..query1.."[SEP]"..query2.."[RESPONSE_TOKEN]"..response
-- Return the input string
return ResponseClassifier(input)
end
function cm.GenerateQuerySearch()
local newdata={}
local contextdb=require(game.ReplicatedStorage.GlobalSpells.ChatbotAlgorithm.Context.Index.FantasyDialogue)
for i,v in testdataset do
newdata[i]={}
local words=cm.splitString(i)
local synomar=cm.GetSynomArray(words,true,false)
task.wait()
for t,o in contextdb do
local Result1,blacklist,score2,weight2,sortedMatches=cm.CompleteQuery(i, o,false,false,true,false,nil,nil,nil,synomar,words)
--cm.CompleteQuery(table.concat(emotiontbl, ""),animdb,1,true,true,false,false,true)
print(sortedMatches)
if sortedMatches~=nil then
for p,o in sortedMatches do
if p<3 then
local score=constructInput("", i,sortedMatches[p].match)
print(score)
local a=sortedMatches[p].address
if newdata[i]==nil then
newdata[i]={}
end
if newdata[i][t]==nil then
newdata[i][t]={}
end
newdata[i][t][p]={score=score,match=sortedMatches[p].match}--p is priority
-- {{[1]=0.5,[2]="a"}, {[1]=0.8,[2]="b"}, {[1]=0.3,[2]="c"}}
--table.sort(newdata[i][t], function(a,b) return a[1] > b[1] end)
end -- table.insert(sortedMatches, {match = bestMatch, count = bestCount, weight = bestweight,truecount=truec,address=i})
end
end
if apicalls>30 then
break
end
Modern large language models use a Transformer architecture. Although given the size of conventional local LLM you would be hard pressed to be able to run it in ROBLOX without GPU acceleration, so building a chatbot using limited computational resources may include a bunch of hacks that separate it from cutting edge methods that brute force the issue using models trained for weeks on end using the most advanced GPUs.
This link is the best resource for training a neural network in roblox, although I’m not sure if it includes a Transformer, Although I do believe it contains the components of a transformer architecture.
DataPredict [Release 1.16] - Machine And Deep Learning Library (Learning AIs, Generative AIs, and more!)
Also, this module is no longer maintained, as I said it was more of a learning experience and it’s a bit convoluted and messy state and extends about 14000 lines of code. But it should be in a usable state given the directions. Hopefully in the future I can find the time and restructure or rewrite the main functionalities of this module. It’s all open sourced and free to use and demonstrates solutions to data science problems in creating chatbots that can solve worded arithmetic, recognize search queries and train using a compressed vocabulary.
It looks really good ngl. It is really impressive and interesting. You usually see chatbots that output predefined responses. But this has real work and It is amazing. The fact that It can be used for an AI as well as a non-ai Chatbot is amazing, truly amazing.
I am an experienced programmer in Roblox and I would LOVE to someday maybe collab with you! We could surely make something good and Big!
I love this project and the effort you are putting on It! Keep Up the good work
1 Like
IF you want to carry on some new work try building a chatbot in ROBLOX you should use this module for the model Training. DataPredict [Release 1.16] - General Machine And Deep Learning Library (Learning AIs, Generative AIs, and more!)
As you train you need a text-encoder to turn the inputs into numbers. This would involve building a vocabulary of the first word received by the cm.GetSynonyms function this function will drastically decrease the complexity of the training data and make your chatbot much more efficient and faster to train by compressing the english language while maintaining coherence.
It would be something similar to
function cm.TrainModel(strings,model)
--local model={}
for i, str in ipairs(strings) do -- loop through the strings in the table
local words=cm.splitString(str)
for t,wo in ipairs(words) do
local prevw,nextw
if wo~="I" then
wo=wo:lower()
end
local s=cm.Getsynonyms(wo,true)
--print(s[1])
if model[s[1]]==nil then
model[s[1]]={}
model[s[1]]["fr"]=1
--model[s[1]]["pw"]={}
model[s[1]]["nw"]={}
-- print(model[s[1]])
end
model[s[1]].fr=model[s[1]].fr+1
if t~=1 then
--local prev=cm.Getsynonyms(words[t-1],true)
--prevw=prev[1]
--if model[s[1]].pw[prevw]==nil and prevw then
---- model[s[1]].pw[prevw]=
-- model[s[1]].pw[prevw]=1
--else model[s[1]].pw[prevw]=model[s[1]].pw[prevw]+1
--end
if t~=#words then
local nex=cm.Getsynonyms(words[t+1],true)
nextw=nex[1]
if model[s[1]].nw[nextw]==nil then model[s[1]].nw[nextw]=1
else model[s[1]].nw[nextw]=model[s[1]].nw[nextw]+1
end
end
end
end
end
As you can see the model in that example indexes the words as a key to access to the number in the subtables. To train using a machine learning model you have to turn all of the words into numbers. So you would index the number into a vocabulary whereas the number equals the word. to unpack your model.
and Viola! You can train a Large Language Model by feeding it datasets similar to how you would train a davinci style model. If you need any help setting up a neural network you can ask the creator of the machine learning library for help and he is pretty nice and helpful.
Basically the TrainModel functions and Predict run functions are pretty close to being compatible with that library it’s just a few changes.
This process would make you the first in ROBLOX to ever make a chatbot using traditional machine learning methods such as the ones documented in his library.
I would say the cons are that it’s experimental, it’s been known to take lots of training to make machine learning models learn what they need about language, the main advantage you’d have is the amount of language compression the getsynonyms function provides which could reduce the
learning space complexity of your model exponentially.
I am doing other stuff currently and but I have a newer version of this module is currently pretty similar but a bit cleaned up and performant. but for the getsynonyms function it’s pretty simple to just call it. The dataset has been thoroughly tested to maintain coherence and meaning of a sentence.
This video demonstrates some of the process, where he creates a chatbot trained on only texts from shakespeare.
To train a chatbot that takes in queries, answers, and a system message you would train it on data structured like that and provide the end user the answer.
You can find conversational datasets in this structure on huggingface.co
My project was more about training a classifier and determining the accuracy based off how good of a response is given.
Similar to something like this
local function ResponseClassifier(payload)
-- Define the API URL and the authorization header
local API_URL = "https://api-inference.huggingface.co/models/tinkoff-ai/response-quality-classifier-tiny"
local headers = {Authorization = Bearkey }
-- Define a function to query the API with a payload
local function query(payload)
-- Encode the payload as a JSON string
local jsonPayload = HttpService:JSONEncode(payload)
-- Send a POST request to the API and get the response
local response = HttpService:PostAsync(API_URL, jsonPayload, Enum.HttpContentType.ApplicationJson, false, headers)
-- Decode the response as a JSON table
local jsonResponse = HttpService:JSONDecode(response)
-- Return the response table
return jsonResponse
end
-- Call the query function with pcall and handle the error
local status, result = pcall(query, payload)
if status then
-- The query function returned successfully, result is the response table
print(result)
else
-- The query function raised an error, result is the error message
result=0
print("Error: " .. result)
end
return result
end
which can be used to train or fine tune a model. You can also use chatGPT for supervised training by having it rate the answers outputs using a sort of GPT style function interaction tool.
Thank you! I will make sure to use it. Here is the current tokenizer:
local NLPM = {}
local separators = {
"·",
",",
".",
";",
":",
"!",
"?",
"-",
"(",
")",
"[",
"]",
"{",
"}",
"<",
">",
"'",
'"',
"/",
"\\",
"|",
"_",
"*",
"&",
"^",
"%",
"$",
"#",
"@",
"`",
"~",
"=",
"+",
"-",
"*",
"/",
"\\",
"|",
"_",
":",
";",
",",
".",
"?",
"!",
"\t",
"\n",
"\r",
"\f",
"\v",
"1",
"2",
"3",
"4",
"5",
"6",
"7",
"8",
"9",
"0",
"ª",
"º",
"...",
"—",
"–",
"‘",
"’",
"“",
"”",
" ",
}
function has(separators,char)
for _,value in ipairs(separators) do
if value == char then
return true
end
end
return false
end
function NLPM.word_tokenize(input:string)
local tokens = {}
local currentToken = ""
for i = 1, #input do
local char = input:sub(i, i)
if has(separators, char) then
if currentToken ~= "" then
currentToken = string.lower(currentToken)
table.insert(tokens, currentToken)
currentToken = ""
end
if char == " " and #tokens > 0 and #currentToken == 0 then
-- Unir espacios con la palabra anterior
tokens[#tokens] = tokens[#tokens] .. char
else
table.insert(tokens, char)
end
else
currentToken = string.lower(currentToken) .. char
end
end
if currentToken ~= "" then
table.insert(tokens, string.lower(currentToken))
end
--print(tokens)
return tokens
end
function NLPM.untokenize(tokens)
return table.concat(tokens)
end
function NLPM.extractWord_fromTokens(tokens,word)
for index,token in pairs(tokens) do
if string.lower(token) == string.lower(word) then
return token
end
end
return nil
end
function NLPM.extractWord_fromInput(input,word)
for i = 1,#input do
local token = string.sub(input,i,i)
if string.lower(token) == string.lower(word) then
return token
end
end
return nil
end
function NLPM.extractNumbers_fromInput(input:string)
local number_pattern = "%d+"
local numbers = {}
for match in input:gmatch(number_pattern) do
table.insert(numbers,match)
end
return numbers
end
function NLPM.extractFirstNumber_fromInput(input:string)
local number_pattern = ("%d+")
local numbers = {}
if input:match(number_pattern) then
table.insert(numbers,input:match(number_pattern))
end
print("EJWE",numbers,"EJWE2")
if #numbers >= 1 then
return tonumber(numbers[1])
end
return nil
end
function NLPM.extractLastNumber_fromInput(input:string)
local number_pattern = "%d+"
local numbers = {}
for match in input:gmatch(number_pattern) do
table.insert(numbers,match)
end
local number = numbers[#numbers]
return tonumber(number)
end
return NLPM
1 Like
That looks good! Chatbots that respond to queries have structured data often encoded in Json to seperate example query and response, for inference inject user query and do next token prediction. The features you extract from training text are inputs, these can be word position, previous word current word next word etc. These inputs can be encoded into matrixes (numbers). and used to train a model, provided the right setup. Using the synonyms=cm.Getsynonyms(word,true) to return the array of synonyms you get compressed language inputs be hashing synonyms[1] or synonyms[#synonyms] this would reduce the complexity of what the model has to learn to chat.
DataPredict [Release 1.16] - General Machine And Deep Learning Library (Learning AIs, Generative AIs, and more!)
Your seperators should be organized
local newseperators={}
local inputnumber=0
for i,token in separators do
inputnumber+=1
newseperators[token]={i,string.len(token)}-Begin to store features of that input.
end
print(newseperators)
function Model(inputtext)
local input={}
local wordarray=cm.splitString(inputtext)--copy and past this function to tokenize inputs.
for i,word in wordarray do
local compresswords=cm.Getsyonyms(word,true)
local vocabelement=compresswords[1]
if newseperators[vocabelement]==nil then
inputnumber+=1{1}
newseperators[vocabelement]={inputnumber,string.len(vocabelement),1}--one is repetition
else
newseperators[vocabelement][3]+=1
end
--etc
end
end
Thank you. I Will make sure to credit you for helping. 
1 Like
Here is how I plan on making the Neural Network
• a table with numerical data will be feed into the NeuralNetwork
• the table Will be iterated thru and you Will be able to choose between using Sigmoid and ReLu
• The weights Will be applied with the bias, and the vector Will be fed to the function you chose
• The final vector Will be added to a table…
1 Like
It Will have 4 hidden layers with 100 neurons each. There Will be 500 weights and 500 bias
The max input length Will be 500
1 Like
I will probably need to import the weights cuz if I do the training in Roblox, my computer bouta blow
1 Like
Another interesting thing I found is this video, which demonstrates creating a GPT-2 like model. Let’s reproduce GPT-2 (124M) - YouTube
It is interesting! Altho I I’ll be dead by the time the video ends 