Mini-GPT Project

Tue, 01 Jul 2025 00:00:00 +0000

Step by Step to a mini-GPT

Trying to create a small mini-version similar to the chatGPT by the end of this project.

Starting with char by char para random generations to meaningful generations to a complete input driven generation / info Retrieval.

01 - Data Preprocessing

Dataset : WikiTex Dataset

cleaned and processed dataset to generate new paragraphs

02 - Bigram Char based model

Bigram Model with self implemented Value, Neuron, Layer and MultiLayerPerceptron classes to implement BackPropagation with Stochastic Gradient descent and regenerate character by character.

Sample output Starting with Random character :

1

grte wouat 1 f Ithe t Ch . ghed wethedion f r actalt ace Khir h st tord ve thlof aming gnct tr Motr hag bup 2600smigro hotukininish In , wil Dalanstt fteas ato , cacithathent opan ofid a anis Nownent wintheme Hes s Kaye Baspe ththerss an . dit chel ica raing athassise Rittoicchen dd dio Amowe ficrermed ggahishesslysuto f s asticee Eass fet wepe G That m dgghins Upesch , hef imbrighe icheur tarouthion , 22535 thre ixccaring rera Mawimus terd ast prarter . hivedeaiured r n , sdutogarid we atonznng

Sample output Starting with “Telugu " :

1

Telugu . ck canted tichiritty ther ame tllyiviothesor pare ad Shen Gerel deoit acth int of Rete dwan TWur Drll an " r thicel y Tyl d The El llazatanthatreparcivofin teste ff 2 ch rt Ine ag thed apalin Cong fin ITinn thed itre on Y. letstenn byt belisselly lorve . is Yoo d Criom Cot Toudes 's Thealo n . , Mat . 'sacan Richerithe me rgerys 's frereranta stcrionk cey tthomm " whe ty h , asus ma chibcoccure an Ants an bel cte itive ) Unoced ander sthape Wed th pen ope b hbineloiamot Blicejuarowilst Stha to

03 - creating new names from existing Indian Names

Loss function used

maximum (log) likelihood estimation

$$ \textrm{L}(x) = \frac{1}{n} \sum_{i = 1}^{n} - log( f(x) ) $$

Loss Report

Model	Context (#Char)	Train Loss	Test Loss

Bigram (Probablistic) Model	1	2.26088

Single Layer NN	1	2.38656
MLP-1hLayer-100-2D	2	2.27822
MLP-1hLayer-300-2D	2	2.09353
MLP-1hLayer-100-3D	2	1.90827
MLP-1hLayer-100-10D	2	1.61364	1.64191
MLP-1hLayer-100-10D (+ softmax - init W2 fix)	2	1.60924	1.64042
MLP-1hLayer-100-10D (+ tanh - Kamming W1 fix)	2	1.58817	1.62057

MLP-1hLayer-100-10D	2	1.59163	1.60631
MLP-2hLayer-100-10D	2	1.53046	1.57615
MLP-2hLayer-100-10D	3	1.39479	1.44245
MLP-3hLayer-100-10D	3	1.35602	1.42019
+ Batch Normalisation
MLP-2hLayer-100-10D	2	1.60034	1.59963
MLP-2hLayer-100-10D	8	1.35103	1.42776
MLP-3hLayer-100-10D	8	1.32197	1.40935
WaveNet + BN
WaveNet-3hLayer-64-10D	8	1.34357	1.41593

best model output till now

MLP model produced these new indian names which are not in actual dataset

Model :

Train Loss : 1.32197 | Test Loss : 1.40935
No of Parameters : 31897
No of hidden layers : 3
No of dimensions used to encode : 10
trained for : 100,000 iterations (with batch_size of 32)
Output :

 1
 2
 3
 4
 5
 6
 7
 8
 9
10


nikkika
bhatham
sand
niik
amat
binti
yogke
narsijal
devy
vens

04 - creating chatgpt with introducing context

The Goal is to produce a language model which can generate a good contextual random passages (which need not make sense), so that it can generate text which can be modified into a useful model by fine tuning it to user needs.

(Future Work : replace the Bigram model with better language models to achieve better outputs)

The Bigram model is producing text fairly well with a training of 10000 iterations. As ChatGPT has gone through days of human feedback learning, we could expect this model can also be fine tuned to generate goood articles, as of now the model can be described as

Loss