LLM Based Embedded Value Trained ChatBot Framework for

Personalized Academic Learning

Chandan Satwani

, Vrashabh Patil, Kavya Morab, Santosh Pattar

and Prema T. Akkasaligar

Department of Computer Science and Engineering, KLE Technological University’s Dr. M. S. Sheshgiri College of

Engineering and Technology, Belagavi, India

Keywords:

LLM-based ChatBot, Personalized Academic Learning, Contextually Relevant Responses, Falcon-7B-Instruct

Model, Data Acquisition and Processing.

Abstract:

Latest developments in natural language processing and machine learning techniques have enabled the de-

velopment of chatbots. However, these chatbots are generalized and not tuned for speciﬁc requirements of

a particular domain. In an academic setting, both learners and teachers today require a specialized chatbot

for their personalized teaching learning experience. In this regard, we propose a novel approach using Large

Language Model (LLM), ﬁne-tuned with embedded value training. It leverages contextual embeddings and

semantic representation to provide tailored educational content. The experimental results demonstrate an im-

provement of 10%, 20%, and 30% for BERT F1, ROUGE, and BLEU scores respectively, when compared to

generic ChatGPT 3.5 and Gemini AI chat applications. These results suggests effectiveness in use of academic

speciﬁc chatbot in improving student engagement, comprehension and retention though personalized learning

experience.

1 INTRODUCTION

In today’s digital age, educational institutions are

rapidly integrating technology into their teaching-

learning process to enhance learning experiences.

Among these technologies, chatbots powered by

Large Language Models (LLMs) are gaining promi-

nence(Yan et al., 2024). These chatbots are designed

to assist in various educational activities, aiming pro-

vide immediate and precise responses to student and

faculty, thereby improving the overall efﬁciency of

academic interactions(Firth et al., 2020).

Various models and technologies have been pro-

posed in the past to enhance chatbot capabilities. For

instance, Montagna et al.(Montagna et al., 2023) in-

troduced Paperplain, a tool aiding medical profes-

sionals in extracting relevant information from clin-

ical papers using Named Entity Recognition (NER)

models. Similarly, Dramatron (Mirowski et al., 2023)

employs hierarchical story generation using the Chin-

chilla LLM, showcasing the potential of large lan-

https://orcid.org/0009-0003-7357-2894

https://orcid.org/0000-0001-9029-5161

https://orcid.org/0000-0002-2214-9389

guage models in creative writing tasks. In health-

care, Omoregbe et al. (Omoregbe et al., 2020) devel-

oped a text-based medical diagnosis system utilizing

NLP and fuzzy logic, demonstrating the adaptabil-

ity of these technologies in diverse ﬁelds. These ex-

isting solutions highlight the versatility and potential

of LLM-based chatbots in addressing domain-speciﬁc

challenges.

Despite the advancements, several challenges per-

sist in the implementation of LLM-based chatbots.

One major issue is maintaining context over pro-

longed interactions, that leads to inaccurate or irrel-

evant responses(Florindi et al., 2024). Additionally,

the performance of these chatbots heavily relies on

the quality and diversity of their training data, that

is often limited. These problems necessitate the de-

velopment of more reliable and focused chatbot solu-

tions tailored to speciﬁc domains, such as education

(

Sar

cevi

c et al., 2024).

To address these challenges, the proposed solution

involves developing a chatbot speciﬁcally trained on

the speciﬁc university syllabus, ensuring that it caters

to the unique needs of students and faculties(Chen

et al., 2024). The methodology includes gathering ex-

tensive course material from textbooks, student notes,

Satwani, C., Patil, V., Morab, K., Pattar, S. and Akkasaligar, P. T.

LLM Based Embedded Value Trained ChatBot Framework for Personalized Academic Learning.

DOI: 10.5220/0013599800004664

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 3rd International Conference on Futuristic Technology (INCOFT 2025) - Volume 2, pages 669-675

ISBN: 978-989-758-763-4

669

and online resources, followed by training the LLM,

creating a model capable of understanding and re-

sponding to academic queries accurately. The chatbot

interface serves as the application layer, facilitating

users to pose questions and obtain exact answers uti-

lizing the trained model(Ooi et al., 2023).

The experimental results show remarkable results.

The accuracy of responses is on par with, other

renowned generative AI models such as ChatGPT

3.5(OpenAI, 2023) and Gemini AI(Google, 2023).

Various testing metrics indicate a signiﬁcant perfor-

mance boost, with the chatbot effectively resolving

queries from both the students and faculties. This

success demonstrates the practical applicability of the

proposed model and its potential to enhance academic

interactions signiﬁcantly. In this regards, the contri-

bution are as follows.

• Domain-Speciﬁc Training: The chatbot is specif-

ically trained on a speciﬁc university data, ensur-

ing high relevance and accuracy of responses.

• Performance and Accuracy: The model ex-

hibits superior performance compared to tradi-

tional models, as evidenced by various testing

metrics.

The rest of the paper is organized as follows. Sec-

tion 2 introduces the problem statement and objec-

tives, followed by the proposed methodology in Sec-

tion 3. Section 4 discusses the implementations and

results obtained. Finally, the paper concludes in Sec-

tion 5.

2 PROBLEM STATEMENT

This section describes our problem statement and ob-

jectives of our work.

A. Problem Statement

Our goal is to design a university-speciﬁc chatbot

that provides detailed, accurate responses to queries.

Users interact with an interface, submitting questions

that are processed by a value-trained model. This

model, trained on diverse academic materials, gener-

ates precise, contextually relevant answers, ensuring

immediate and effective support for students and fac-

ulty.

B. Objectives

The chatbot is designed for university students

and faculty, addressing limitations such as maintain-

ing context over long interactions, data quality, and

privacy concerns. The objectives are:

1. Support Stakeholder Queries: The chatbot aims

to address stakeholder queries efﬁciently, elimi-

nating the need for manual searches.

2. Provide Unambiguous and Precise Responses:

The chatbot must deliver swift, accurate, and clear

answers, enhancing reliability and efﬁciency in

academic support.

3 SYSTEM MODEL

This section provides a comprehensive overview of

the architecture and operational workﬂow of the pro-

posed chatbot as shown Fig 1. It starts by explain-

ing the data collec- tion process, particularly how the

training data is curated in alignment with the univer-

sity’s syllabus. It then outlines the embedding model

used to convert this data into a structured format suit-

able for model training. Further, the model train-

ing phase is described, detailing the development and

ﬁne-tuning of the machine learning model.

Training Data

Embedding Model

Embedded Data

Model Training

Model

ChatBot

Response

User

Query

Figure 1: Flowchart of the proposed Chatbot System.

3.1 Model Description

The workﬂow begins with the acquisition of training

data, speciﬁcally collected in alignment with the uni-

versity’s syl- labus. This data is processed by the em-

bedding model, convert- ing it into structured embed-

ded data. The embedded data then undergoes model

training, resulting in a ﬁnely tuned machine learning

model. Once the model is integrated into the chat-

bot system, users input their queries, and the chat-

bot leverages the trained model to promptly generate

and deliver accurate results. This process ensures the

chatbot to provide precise, unambiguous, and timely

responses, signiﬁcantly enhancing the academic sup-

port experience for all stakeholders.

INCOFT 2025 - International Conference on Futuristic Technology

670

3.2 Data Description

The dataset comprises a diverse collection of aca-

demic materials relevant to courses at a speciﬁc uni-

versity. It includes textbooks, student notes, and on-

line resources, primarily in PDF format. The dataset

is carefully curated to ensure com- prehensive cover-

age of the course content, providing a rich source of

information for embedding and subsequent use in the

chatbot framework. Detailed descriptions and statis-

tics of the dataset are presented in the Table1.

Table 1: Details of the Dataset.

Courses Type Materials Size

Course–1 Theoretical Notes + Text Book 2048 pages

Course–2 Theoretical Online Resources + Notes 1980 pages

Course–3 Theoretical Online Resources + Notes 1880 pages

Course–4 Laboratory Notes + Online Resources 1500 pages

3.3 Embedding

The embedding process involves several stages to en-

sure the seamless integration of the instructor mod-

ule within the LLM-based chatbot framework, tai-

lored for a speciﬁc university. Initially, a compre-

hensive data collection phase is undertaken, gather-

ing information from a variety of academic sources.

These sources include textbooks, student notes, on-

line materials, and any additional relevant resources

related to speciﬁc courses. This data, often in PDF

format, undergoes an extraction process to isolate the

pertinent content required for embedding. Following

data extraction, a crucial preprocessing stage is em-

ployed to cleanse the data. This involves removing

noise, correcting inconsistencies, and standardizing

the text to ensure it is in an optimal format for em-

bedding. This step is vital to maintain the integrity

and quality of the data, enabling effective embedding.

The preprocessed data is then subjected to embedding

algorithm using instructor-xl instructor model, trans-

forming the textual information into numerical repre-

sentations as shown in Table 2. These embeddings

capture the semantic nuances and contextual mean-

ings of the text, making them comprehensible for the

LLM. The instructor module, now enriched with these

sophisticated embeddings, is integrated into the chat-

bot framework. This integration allows the chatbot to

utilize the embedded knowledge, providing person-

alized and contextually accurate academic assistance

to students. The instructor module leverages the em-

bedded data to generate detailed explanations, answer

complex queries, and offer tailored learning support.

This embedding process ensures that the chatbot is

not only informed by a wide array of academic ma-

terials but also capable of delivering precise and rele-

vant educational guidance, enhancing the learning ex-

perience for students at the speciﬁed university.

Table 2: Embedded Dataset Details.

CT T S NC UPC NS WFS

Course–1 Notes 60MB 10 7000–8500 60–65 12–14

Text Book 60MB 12 7000–9000 60–65 12–14

Course–2 Online Res. 60MB 10 7000–8500 60–65 12–14

Notes 60MB 10 7000–8500 60–65 12–14

Course–3 Online Res. 60MB 12 7000–9500 60–65 12–14

Notes 60MB 10 7000–8500 60–65 12–14

Course–4 Notes 60MB 10 7000–8500 60–65 12–14

Online Res. 60MB 10 7000–8500 60–65 12–14

CT: Course Type, T: Type of Material, S: Size of the dataset

NC: Number of chunks generated, UPC: Unique Word Per Chunk

NS: Number of sentences, WFS: Word Frequency Per Sentence

3.4 Training model

The training model serves as a cornerstone of the

chatbot framework. It utilizes the embedded data and

employs an advanced algorithm to generate a model

proﬁciency in pro- cessing user queries and produc-

ing responses. Speciﬁcally, the Falcon-7B-Instruct

model architecture (Almazrouei et al., 2023), an LLM

for text- based chat applications, is leveraged. The

model features 32 layers, a 4544-dimensional hidden

space, 64-dimensional at- tention heads, a vocabulary

of 65024, and supports 2048-token sequences. This

model excels in text generation, prioritizing speed

and efﬁciency over contextual depth, distinguishing

it from other models. To tailor the model to spe-

ciﬁc require- ments, hyperparameters are adjusted,

including temperature hyperparameter, token gener-

ation, and maximum length. This customization en-

sures optimal performance aligned with the speciﬁed

needs. Further, the model undergoes training using

university-speciﬁc data from the embedding module.

Once the training phase is complete, the reﬁned model

is stored and integrated into the chatbot system. The

resultant model, ﬁnely tuned to the speciﬁcations as

shown in Table 3, is thus equipped to handle user

queries with precision and generate accurate, relevant

responses.

LLM Based Embedded Value Trained ChatBot Framework for Personalized Academic Learning

671

Table 3: Falcon - 7B LLM Speciﬁcations.

Hyperparameter Value

Layers 32

d–model 4544

head–dim 64

Vocabulary 65024

Sequence length 2048

Input Text (x)

Instruction (I

)

Concatenation (I

⊕ x)

GTR Encoder

(Base/Large/XL)

Mean Pooling

Task-speciﬁc Embedding E

, x)

(a) Instructor

Input

Embedding Layer

Self-Attention

Feed Forward

Self-Attention

Feed Forward

Self-Attention

Feed Forward

Output

N Layers

(b) GTR Encoder

Figure 2: Instructor-XL Architecture.

Output probabilities

Softmax

Linear

Add & Norm

FFN

Add & Norm

Multi-Head

Attention

Add & Norm

Masked Multi-

Head Attention

Input embedding

Figure 3: Falcon-7B Decoder Architecture.

3.5 Performance Evaluation

Parameters

The following parameters are used to assess the pro-

posed model’s performance.

BERT–F1 Score: BERT–F1 Score is a metric that

evaluates the quality of text generation by comparing

the con- textual embeddings of the generated text and

the reference text using BERT, a powerful language

model. Unlike traditional metrics, BERT–F1 score

captures subtle nuances in meaning and context, pro-

viding a more reﬁned assessment of textual similarity

and quality(Shankar et al., 2024). The equation for

the BERT–F1 score is as follows.

BERT

| ˆy|

∑

ˆy

∈ ˆy

max

∈y

cosine similarity

z}|{

⊤

· y

(1)

BERT

|y|

∑

∈y

max

ˆy

∈ ˆy

cosine similarity

z}|{

⊤

· y

(2)

BERTF

= 2 ·

BERT

· Re

BERT

+ Re

BERT

(3)

where, Pr

BERT

represents the precision score of

BERTScore, while Re

BERT

represents the recall score.

The terms y

and ˆy

are individual contextual embed-

dings. The dot product, y

⊤

· y

, reﬂects the cosine

similarity between embeddings y

and y

ROUGE Score: ROUGE measures the overlap of

n-grams, word sequences, and word pairs between the

generated output and reference text, indicating rele-

vance and completeness(Shankar et al., 2024). The

following equations show the calculation for ROUGE

scores.

ROUGE-1 =

Number of overlapping unigrams

Total number of unigrams in the reference summary

(4)

ROUGE-2 =

Number of overlapping bigrams

Total number of bigrams in the reference summary

(5)

ROUGE-L =

LCS(C, R)

Length of the reference summary

(6)

where LCS(C, R) is the length of the longest common

subsequence between the candidate summary C and

the reference summary R.

INCOFT 2025 - International Conference on Futuristic Technology

672

BLEU Scores: BLEU evaluates the precision of the

generated text by comparing it to one or more refer-

ence texts, focusing on coherence and accuracy. The

BLEU score is calculated using the following equa-

tions:

Precision

Number of n-grams in the candidate that appear in the references

Total number of n-grams in the candidate

(7)

BP =

(

1 if Candidate Length > Reference Length,



1−

Reference Length

Candidate Length



if Candidate Length ≤ Reference Length.

(8)

BLEU = BP × exp

∑

n=1

logPrecision

(9)

The score ranges for BLEU, ROUGE, and BERT

score all span from 0 to 1, with higher scores indicat-

ing better quality: BLEU scores above 0.5 are con-

sidered good for translation quality, ROUGE scores

above 0.6 signify good overlap with reference texts,

and BERT scores above 0.9 reﬂect high contextual

similarity.

4 IMPLEMENTATIONS AND

PERFORMANCE ANALYSIS

The experiments are conducted on Google Colab,

which provides access to CPUs, GPUs like Tesla K80,

T4, P4, and P100, TPUs, up to 12 GB RAM, and in-

tegrated Google Drive storage. The environment uses

libraries such as langchain, torch, transformers, json,

and matplotlib for efﬁcient model training, text pro-

cessing, and data visualization.

Data is collected from various sources, includ-

ing textbooks, notes, and online materials, form-

ing a comprehensive dataset of approximately 700

megabytes. This data is formatted into PDFs and pro-

cessed using the PyPDF2 module in Python to extract

textual information. The extracted text is then fed

into the proposed embedding module, that performs

meaningful embeddings based on provided instruc-

tions, resulting in a context-rich, high-quality training

dataset for the LLM.

The Instructor-XL model as shown in Fig. 2 en-

erates task-speciﬁc embeddings by merging text in-

puts with task instructions, handling multiple tasks

without needing extra training. It adds an extra ﬁne-

tuning layer and processes data in overlapping chunks

to ensure high quality and ac- curate relationships.

The model is trained with General Text Representa-

tion (GTR) architecture for multitask approach on 330

diverse datasets. The model ﬁne-tunes embeddings

with contrastive loss to separate, related from unre-

lated text pairs. It excels in tasks like classiﬁcation,

information retrieval, and semantic similarity. Simi-

larity between texts is determined by cosine similarity

of their embeddings. Higher the similarity indicates

greater textual closeness.

Table 4: Sample Course Queries.

Query Question Answer

Q1 Discuss the con-

cept of operational

ampliﬁers (op-

amps) in analog

electronics?

Op-amps are electronic circuits that am-

plify small signals and convert them into

large signals that can be used within a

device.

Q2 What are key com-

ponents of com-

puter organization

and architecture?

The key components of computer orga-

nization and architecture include proces-

sor (CPU), memory (RAM/ROM), stor-

age devices (hard drives/SSDs), control

unit (CPU’s processor), and input/output

devices (keyboard/input devices).

Q3 How does the con-

cept of eigenvalues

and eigenvectors

relate to Princi-

pal Component

Analysis (PCA) in

statistics and data

analysis?

Eigenvalues determine the variance cap-

tured by each principal component,

while eigenvectors deﬁne the direction

of these components. Eigenvectors, as

principal components, are linear combi-

nations of original data points that cap-

ture the most variance.

Q4 Explain the con-

cept of matrix rank

and its signiﬁcance

in linear algebra?

A matrix rank is used to solve problems

that involve computing the eigenvectors

and eigenvalues of a matrix, and thus,

matrix ranks have played a key role in

the development of modern linear alge-

bra theory and applications.

This embedded dataset as shown in Table 2 is used

to train the customized Falcon-7B model, ﬁne-tuned

to retain the meaning of the text during the embed-

ding process, ensuring accurate and contextually rel-

evant responses. The ﬁnal chatbot system integrates

this well-trained model with an interface. The user

submits the query, to the model to generate pre- cise

and relevant responses, enhancing academic support

for stakeholders at the speciﬁed university. The Table

4 shows the sample query asked to the model by the

stakeholders.

When the user asks a query to the chatbot model,

the chatbot processes the input using an embedding

model. The embedding model converts the text into a

structured format and uses the Instructor-XL model to

generate meaningful embedding. The embedded data

is then used to train the Falcon-7B-Instruct model,

ﬁne-tuned with embedded value training. The trained

model generates a response based on the input query.

The Table 5 hows queries asked to both ChatGPT-

LLM Based Embedded Value Trained ChatBot Framework for Personalized Academic Learning

673

3.5 and the proposed model during testing phase, with

ratings provided by the stakeholders to the responses.

The proposed model demonstrates superior perfor-

mance compared to ChatGPT-3.5, with more accurate

and context-aware responses

Table 5: Course Queries and Ratings

Query Chat-GPT 3.5 Proposed Model

Q1 4.3 4.5

Q2 4.0 4.4

Q3 4.2 4.6

Q4 3.8 4.4

In the chatbot implementation, ﬁne-tuning of the

LLM is performed to optimize response time efﬁ-

ciency. From Fig. 4 we ﬁnd that extensive experimen-

tation and rigorous opti- mization leads to a signiﬁ-

cant reduction in response time. The Fig. 4 The Fig. 4

shows response generation time ranging from 2 to 15

seconds. This accomplishment not only validates the

tuning process but also highlights the model’s capa-

bility to deliver prompt and accurate answers. Thus,

the chatbot demonstrates the practical application and

efﬁciency of the optimized LLM model in real-world

scenarios.

Q1 Q2 Q3 Q4

Queries

Response Generation Time (in seconds)

Figure 4: Response Times of LLM Model.

The BERT F1 score helps ensure query responses

are contextually accurate and meaningful. The Fig.5

shows comparison of BERTF1Score for proposed

method ChatGPT,GeminiAI. From the Fig 5 it is

found that the proposed model outperforms ChatGPT-

3.5 and Gemini AI in machine translation, text sum-

marization, and dialogue generation. This leads to

a signiﬁcant impact on educational chatbots, as the

proposed model can provide more precise, relevant,

and high- ﬁdelity interactions, enhancing personal-

ized learning experi- ences and ensuring students re-

ceive high-quality support

ROUGE and BLEU are evaluation metrics for as-

sessing the quality of generated text in natural lan-

BERT F1 Score

0.2

0.4

0.6

0.8

0.9

0.85

0.82

Scores

Proposed Model ChatGPT3.5 Gemini AI

Figure 5: Comparison of Bert F1 Scores.

ROUGE-1 ROUGE-2 ROUGE-L

0.2

0.4

0.6

0.4

0.54

0.4

0.19

0.32

0.33

0.1

0.25

Scores

Proposed Model ChatGPT 3.5 Gemini AI

Figure 6: Comparison of ROUGE Scores

guage processing. The Fig.6 shows that the pro-

posed model achieves parity with GPT-3.5 and out-

performs Gemini AI by 30% due to its exceptional

ability to generate contextually relevant and compre-

hensive responses, as reﬂected in the high ROUGE

scores as shown in Fig. 6. The high BLEU scores

further highlight the model’s precision and ﬂuency

in language generation as shown in Fig. 7. his su-

perior performance enhances the effectiveness of ed-

ucational chatbots by ensuring they provide coher-

ent, accu- rate, and contextually appropriate support,

thereby improving the overall learning experience for

students.

The results obtained are a direct outcome of the

modiﬁ- cations implemented. These enhancements

give the proposed model an edge over others, making

it unique and poised for great success in the future.

5 CONCLUSIONS

The integration of chatbots in education presents

a signiﬁ- cant opportunity to enhance personalized

learning experiences for students. However, existing

language models and chatbots often face challenges

in accurately understanding and respond- ing to the

INCOFT 2025 - International Conference on Futuristic Technology

674

BLEU scores

0.1

0.2

0.3

0.4

0.3

0.11

7 · 10

−2

Scores

Proposed Model ChatGPT 3.5 Gemini AI

Figure 7: Comparison of BLEU Scores.

nuanced needs of learners, leading to inconsistencies

and gaps in educational delivery. The proposed ap-

proach leverages an LLM ﬁne-tuned with embedded

value training, addresses these challenges by ensur-

ing more contextually rele- vant and comprehensive

responses. The superior performance of the model, as

demonstrated by high ROUGE and BLEU scores, in-

dicates its proﬁciency in generating coherent and pre-

cise language, surpassing both GPT-3.5 and Gemini

AI. These results underscore the model’s potential to

improve educational outcomes by providing accurate

and meaningful interactions. Future work will focus

on further reﬁning the model to enhance its adaptabil-

ity and scalability, ensuring it can cater to a diverse

range of educational contexts and needs. This con-

tinued development aims to solidify the role of ad-

vanced chatbots in shaping the future of personalized

academic learning.

REFERENCES

Almazrouei, E., Alobeidli, H., Alshamsi, A., Cappelli, A.,

Cojocaru, R., Debbah, M., Gofﬁnet,

E., Hesslow,

D., Launay, J., Malartic, Q., et al. (2023). The fal-

con series of open language models. arXiv preprint

arXiv:2311.16867.

Chen, J., Lu, X., Du, Y., Rejtig, M., Bagley, R., Horn, M.,

and Wilensky, U. (2024). Learning agent-based mod-

eling with llm companions: Experiences of novices

and experts using chatgpt & netlogo chat. In Proceed-

ings of the CHI Conf. on Human Factors in Computing

Systems, pages 1–18. ACM.

Firth, J. A., Torous, J., and Firth, J. (2020). Exploring the

impact of internet use on memory and attention pro-

cesses. International Journal of Environmental Re-

search and Public Health, 17(24).

Florindi, F., Fedele, P., and Dimitri, G. M. (2024). A novel

solution for the development of a sentimental analysis

chatbot integrating chatgpt. Personal and Ubiquitous

Computing, pages 1–14.

Google (2023). Gemini AI. https://gemini.google.com/.

Text-based AI Model.

Mirowski, P., Mathewson, K. W., Pittman, J., and Evans,

R. (2023). Co-writing screenplays and theatre scripts

with language models: Evaluation by industry profes-

sionals. In Proceedings of the 2023 CHI Conference

on Human Factors in Computing Systems. ACM.

Montagna, S., Ferretti, S., Klopfenstein, L. C., Florio, A.,

and Pengo, M. F. (2023). Data decentralisation of

llm-based chatbot systems in chronic disease self-

management. In Proceedings of the 2023 ACM Con-

ference on Information Technology for Social Good,

pages 205–212.

Omoregbe, N. A. I., Ndaman, I. O., Misra, S., Abayomi-

Alli, O. O., Dama

sevi

cius, R., and Dogra, A. (2020).

Text messaging-based medical diagnosis using natu-

ral language processing and fuzzy logic. Journal of

Healthcare Engineering, 2020:1–14.

Ooi, K.-B., Tan, G. W.-H., Al-Emran, M., Al-Sharaﬁ,

M. A., Capatina, A., Chakraborty, A., Dwivedi, Y. K.,

Huang, T.-L., Kar, A. K., Lee, V.-H., et al. (2023). The

potential of generative artiﬁcial intelligence across

disciplines: Perspectives and future directions. Jour-

nal of Computer Information Systems, pages 1–32.

OpenAI (2023). ChatGPT (v3.5). https://chat.openai.com.

Large Language Model.

Shankar, S., Zamﬁrescu-Pereira, J., Hartmann, B.,

Parameswaran, A. G., and Arawjo, I. (2024). Who

validates the validators? aligning llm-assisted evalu-

ation of llm outputs with human preferences. arXiv

preprint arXiv:2404.12272.

Sar

cevi

c, A., Tomi

c, I., Merlin, A., and Horvat, M.

(2024). Enhancing programming education with

open-source generative ai chatbots. In 2024 47th

MIPRO ICT and Electronics Convention (MIPRO),

pages 2051–2056.

Yan, L., Sha, L., Zhao, L., Li, Y., Martinez-Maldonado, R.,

Chen, G., Li, X., Jin, Y., and Ga

sevi

c, D. (2024). Prac-

tical and ethical challenges of large language models

in education: A systematic scoping review. British

Journal of Educational Technology, 55(1):90–112.

LLM Based Embedded Value Trained ChatBot Framework for Personalized Academic Learning

675