Case Studies assignment On Artificial Intelligence Based Program Tools

1. Ethical Principles for Large Language Models

Large language models (LLMs) hold significant potential for various applications but
also pose ethical risks. Developers and organizations must adopt several ethical
principles to ensure responsible deployment. These principles include:

● Fairness: Models should be designed to avoid discrimination and bias against

certain groups.
● Transparency: Organizations should disclose how models are trained and
make the decision-making processes clear.
● Accountability: Developers must take responsibility for how their models are
used and the consequences of their output.
● Privacy: Sensitive data used for training should be handled with the highest
level of protection to maintain privacy.

These principles help guide decision-making processes by ensuring models are not
used in ways that could harm individuals or society, particularly in sectors such as
healthcare and finance, where fairness and privacy are crucial.

Example: In the healthcare industry, a medical model should ensure transparency in

how decisions are made, giving patients and doctors insight into how an AI diagnosis
was generated, thereby building trust.

2. Risks of Large Language Models

While large language models show impressive capabilities in generating human-like

text, they bring several risks that developers must address to avoid potential misuse:

● Misinformation: LLMs can inadvertently generate false or misleading

information, especially if trained on unreliable data sources.
● Bias: LLMs often replicate societal biases present in the training data, resulting
in biased outputs.
● Security Concerns: Malicious actors could exploit these models to create
 harmful content, such as deep fakes or spam.

Developers can create safeguards by implementing real-time content moderation,

conducting regular audits for bias, and introducing ethical guidelines to monitor
usage. Ensuring that these models are regularly updated with fact-checked data can
also minimize the spread of misinformation.

Example: In the news industry, AI-generated content can be used to automatically

write articles. However, without proper monitoring and content filtering, this could
result in the distribution of misleading or false information.

3. Ethical Dilemmas in AI Integration

AI integration across fields such as education, healthcare, and customer service

presents various ethical challenges. These dilemmas include:

● Job Displacement: The widespread use of AI could displace human workers,

especially in routine tasks.
● Over-reliance on AI: There is a risk that individuals or organizations may rely
too heavily on AI systems, reducing human oversight.
● Bias in Decision-Making: If AI models are not properly trained or validated,
they may introduce biases in crucial decisions, like hiring or legal judgments.

To strike a balance, organizations should focus on using AI to augment human

abilities rather than replace them. In customer service, for example, AI can handle
routine inquiries, but complex or sensitive cases should involve human agents for
better judgment and empathy.

Example: In the healthcare sector, while AI could assist with diagnostics, human
doctors should still make the final decisions, ensuring that AI supports rather than
overrides professional expertise.

4. Challenges in Training Large Language Models

Training large language models requires considerable computational resources, data,

and time. The main challenges include:

● High Computational Costs: Training these models requires massive

computational power, which can be costly and environmentally taxing.
● Data Quality and Availability: Obtaining diverse, unbiased, and high-quality
datasets is crucial for training robust models.
● Energy Consumption: The energy consumption of large-scale training is a
major concern, given its environmental impact.

To address these challenges, organizations can employ techniques like model

distillation (which reduces model size without losing performance) and transfer
learning (using pre-trained models to reduce the need for vast datasets and
computational resources).

Example: OpenAI’s GPT-3, for instance, has an enormous environmental footprint

due to its size and training data. To mitigate this, OpenAI could use energy-efficient
models or employ transfer learning to fine-tune smaller models for specific tasks.

5. Guidelines for Minimizing Harm in Model Deployment

To minimize harm when deploying AI models, organizations can adopt a set of

practices:

● Data Privacy: Ensure that models are trained on secure, anonymized data to
protect user privacy.
● Bias Mitigation: Implement continuous testing and audits to detect and address
bias in AI models.
● Security Measures: Enforce strong security protocols to safeguard models
against exploitation or misuse.
● Ethical Review: Conduct regular ethical reviews to assess the impact of AI
deployment on individuals and communities.

Consistent application of these standards across different industries (such as

healthcare, finance, and education) will help ensure responsible AI deployment and
reduce risks like discrimination and data breaches.

Example: A financial institution using AI to assess creditworthiness must ensure that

the model is transparent, secure, and free of biases that could unfairly disadvantage
certain groups of people.

6. Origins and Manifestation of Bias in AI Models

Bias in AI models often stems from the data used during training. If the training
dataset reflects societal biases or underrepresents certain groups, the model will likely
exhibit biased behavior. This bias can manifest in various ways:

● Stereotyping: Models may generate outputs that reinforce negative

stereotypes, such as associating certain words or topics with specific groups.
● Discriminatory Outputs: A model trained on biased data may produce outputs
that unfairly favor or disadvantage certain demographic groups.

Platforms like Hugging Face’s Transformers can exacerbate bias if not carefully
curated. Developers need to be aware of the data’s origins and actively work to ensure
diversity and balance in the training sets.
Example: In hiring tools, an AI trained on data from predominantly male employees
might show a preference for male candidates, perpetuating gender inequality.

7. Bias in NLP Tasks

In NLP tasks such as sentiment analysis, translation, and text generation, bias can
lead to significant issues:

● Sentiment Analysis: A model might misinterpret the sentiment of a text based

on the writer’s gender or ethnicity.
● Language Translation: Biases in translation models could lead to inaccurate
translations, especially in the context of gendered language.
● Text Generation: Biased AI-generated content could inadvertently promote
harmful stereotypes or misinformation.

These biased outputs can affect industries like hiring, legal assessments, or customer
service. For instance, a biased sentiment analysis tool might unfairly evaluate
candidates in a recruitment process, leading to discrimination.

Example: A customer service AI might show bias in responding to customers from

different demographics, potentially causing frustration and dissatisfaction.

8. Ethical Concerns in Sensitive Fields

In domains like healthcare or criminal justice, biased AI models pose significant

risks:

● Healthcare: Biased medical models could result in incorrect diagnoses or

treatment recommendations for certain populations, undermining trust in
AI-powered healthcare.
● Criminal Justice: If an AI system used in legal assessments is biased, it could
lead to unjust outcomes, such as higher bail amounts or prison sentences for
minority groups.

Ethical concerns arise when AI models influence decisions that affect people’s lives,
such as determining criminal sentencing or medical diagnoses. Organizations must
ensure fairness and accuracy to maintain public trust.

Example: In criminal justice, biased risk assessment tools might unfairly increase the
likelihood of parole denial for minority inmates, perpetuating systemic inequality.

9. Mitigating Bias in AI Models

Reducing bias in AI models is essential, especially in high-stakes fields like

healthcare and criminal justice. Effective methods to mitigate bias include:

● Data Audits: Regularly auditing training data to identify and correct biased
samples.
● Debiasing Algorithms: Implementing algorithms designed to adjust for bias
during model training.
● Rigorous Testing: Continuously testing models on diverse datasets to detect
biased outcomes and refine their behavior.

Integrating these methods into workflows ensures that AI models used in sensitive
fields are both fair and reliable.

Example: In healthcare, debiasing algorithms could help prevent medical AI systems

from giving less accurate diagnoses for underrepresented populations.

10. Hugging Face Transformers Library for NLP Tasks

The Hugging Face Transformers library is a popular tool for using pre-trained NLP
models. Key architectural elements include:

● Tokenizers: These break down text into manageable units (tokens), ensuring
that the model can process text effectively.
● Model Configurations: These contain the architecture details that define how
the model processes data.
● Pipelines: Hugging Face offers pre-built pipelines for tasks like text
classification, translation, and summarization, which simplify model
deployment.

These components make it easy to load, customize, and deploy complex models,
streamlining the NLP process for various use cases.

Example: Hugging Face’s tokenizer allows a user to preprocess text efficiently for
model input, while pipelines offer ready-to-use models for tasks like question
answering, saving time in developing custom solutions.

11. Loading and Fine-Tuning Pre-Trained Models

Fine-tuning a pre-trained model involves several steps:

● Selecting the Right Model: Choose a model suitable for the task (e.g., BERT
for classification).
● Data Preparation: Preprocess data, including tokenization and formatting.
● Parameter Adjustment: Tune hyperparameters to optimize performance.

This multi-step process allows you to adapt a pre-trained model to your specific NLP
task while achieving high performance with minimal resources
12. Query Expansion in Information Retrieval

Query expansion is a technique used to enhance the accuracy of information retrieval

by broadening or refining user queries, making them more effective in retrieving
relevant results. This helps overcome limitations such as ambiguity or lack of
specificity in user queries.

● How it Works: Query expansion involves adding related terms, synonyms, or

semantically similar phrases to a user's original query to improve the likelihood
of retrieving relevant results.
● Advanced Language Models for Query Expansion: Models like BERT or
GPT can help expand queries by suggesting semantically related terms or
alternative ways of phrasing a search query. These models understand the
context and can provide more relevant suggestions than traditional
keyword-based systems.

Example: A search engine might expand the query “smartphone repair” by adding
related terms like “mobile phone fix” or “phone troubleshooting,” improving the
chances of retrieving more relevant results.

Benefits:

● Better Retrieval Results: By expanding queries, the system can understand the
intent behind the query more effectively, leading to more relevant results.
● Improved User Experience: It helps users who may not know the precise
terminology to use or who are unsure of how to phrase their search queries.

13. Challenges in Low-Resource Languages for Unstructured Text Extraction

In low-resource languages, extracting information from unstructured text is

particularly challenging due to several factors:

● Lack of Annotated Datasets: Many low-resource languages lack large,

annotated datasets needed for training NLP models.
● Limited Linguistic Resources: There are fewer NLP tools, dictionaries, and
resources for low-resource languages.

Transfer Learning can be helpful in addressing these challenges by leveraging

pre-trained models from resource-rich languages (like English or Chinese) to improve
performance in low-resource languages.

● Fine-tuning: Pre-trained models can be adapted (fine-tuned) on a small set of

data from the target language.
● Zero-Shot Learning: This technique involves applying a model trained on a
resource-rich language to tasks in low-resource languages without additional
 fine-tuning, based on generalizable features learned from similar languages.

Example: A BERT model trained on English text could be fine-tuned to work on a

low-resource language like Swahili, using transfer learning techniques to bridge the
gap.

14. Machine Translation: RBMT, SMT, and NMT

Machine translation systems have evolved significantly over time:

● Rule-Based Machine Translation (RBMT): Early systems used predefined

linguistic rules and dictionaries to translate text. It was effective when the
languages had clear grammatical rules and structures but limited in handling
ambiguities or context.
Strengths: Precise and works well with specific, rule-driven languages.
Limitations: Scalability issues and struggles with idiomatic expressions or
language variations.
● Statistical Machine Translation (SMT): Based on statistical analysis of
bilingual corpora, SMT uses probabilities to determine the best translation. It’s
more flexible than RBMT but still struggles with fluency and capturing
contextual meaning.
Strengths: Can handle a wider range of language pairs and nuances.
Limitations: It requires large datasets and often produces translations that
sound less natural or fluid.
● Neural Machine Translation (NMT): Uses deep learning models, particularly
transformers, to learn entire sentence structures and context. NMT has become
the state-of-the-art method, offering significant improvements in the fluency
and naturalness of translations.
Strengths: High-quality, contextually relevant translations with
natural-sounding results.
Limitations: Requires vast computational resources for training.

Example: Google Translate’s shift from SMT to NMT drastically improved

translation quality, especially in languages with complex syntactic structures, such as
Chinese or Arabic.

15. Information Extraction (IE) Steps

Information Extraction (IE) is a process of automatically identifying and extracting

structured data from unstructured text, which is crucial for converting raw text into
usable, actionable information. The typical steps in IE include:

● Named Entity Recognition (NER): This step identifies proper nouns in the
text, such as names of people, places, dates, and organizations.
● Relation Extraction: This step finds the relationships between identified
 entities, such as “works at,” “located in,” or “born on.”
● Event Detection: This involves identifying specific events or actions described
in the text and linking them to the relevant entities involved.

Example: From a news article about a politician’s visit to a foreign country, NER
would identify “John Doe” as a person and “Paris” as a location, while Relation
Extraction would link “John Doe” to “visit” and “Paris,” and Event Detection would
highlight the event as a “political visit.”

These steps help turn unstructured data, like news articles, social media posts, or legal
documents, into structured, usable information for further analysis or processing.

16. Challenges in Structuring Data from Unstructured Text

Extracting structured information from unstructured text can be difficult due to the
inherent challenges in natural language:

● Ambiguity: Words and phrases can have multiple meanings depending on

context.
● Contextual Dependencies: Certain meanings only emerge when you consider
the surrounding context.
● Language Variability: Different styles of writing, slang, and dialects can make
it hard to consistently extract structured data.

Examples of Transformer-based Models for IE:

● BERT: Pre-trained on large amounts of text, BERT can be fine-tuned for tasks
like named entity recognition (NER) and relation extraction.
● T5 (Text-to-Text Transfer Transformer): T5 is a versatile transformer that
frames various NLP tasks as a text-to-text problem, including information
extraction tasks.
● RoBERTa: A variant of BERT, RoBERTa has been shown to perform
particularly well in tasks involving text understanding and entity recognition.

These models are trained to understand context better than traditional NLP models,
making them useful for complex IE tasks, such as sentiment analysis and extracting
relationships between entities.

Example: A BERT-based model could be trained to recognize key entities in a

medical document, such as medications, dosages, and treatment procedures, making it
possible to structure information for later use in patient records.

17. Building a Multi-Functional NLP System for a Multinational Organization

Given the requirements for a multinational organization to build a multi-functional

NLP system, the following considerations are crucial:

1. Speech Recognition Models:

○ Wav2Vec 2.0: A powerful pre-trained model from Hugging Face,
designed for automatic speech recognition (ASR). It can handle noisy
speech and provide highly accurate transcriptions.
○ DeepSpeech: Another popular speech-to-text model that can be
integrated into the system to handle voice-based data efficiently.
2. Machine Translation Models:
○ MarianMT: A multilingual machine translation model that supports
multiple language pairs, making it ideal for a multilingual organization.
○ MBart: A transformer model designed for translation tasks, which
supports 25 languages and is useful for handling a variety of language
pairs efficiently.
3. Exploring Pre-Trained Models with Hugging Face:
Hugging Face provides tools like model hubs and pre-trained models that help
quickly access and test models for specific NLP tasks. The user-friendly
interface allows quick exploration of different architectures, like BERT, GPT,
or T5, to match the organization's needs.
4. Transfer Learning and Fine-Tuning:
Transfer learning enables adapting pre-trained models to the organization’s
specific tasks. For example, fine-tuning a BERT model for sentiment analysis
using the company’s proprietary customer service data involves the following:
○ Data Preparation: Prepare data by tokenizing text and converting
labels.
○ Hyperparameter Tuning: Adjust parameters like learning rate, batch
size, etc., to optimize the model.
○ Model Evaluation: Evaluate the model’s performance using metrics like
accuracy, F1 score, etc.
5. Integrating Gemma for Performance Optimization:
Gemma is designed to optimize the training process by providing tools for
hyperparameter tuning, data augmentation, and model evaluation. It can help
fine-tune models for better accuracy and efficiency, particularly for
high-demand tasks.
6. Architecture Overview of Hugging Face Transformers:
The Hugging Face Transformers library offers a unified framework for various
NLP tasks, with easy-to-use APIs and pre-trained models. It includes
components like tokenizers, model configurations, and pipelines, which
simplify the model selection, training, and deployment process for complex
systems across multiple languages.

18. Custom NLP Models for Departments in a Global Tech Company

A global tech company implementing custom NLP models for different departments
(customer service, financial analysis, healthcare support) can achieve optimal
performance by fine-tuning pre-trained models to meet their specific needs.

● Customer Service: Use models like BERT or DistilBERT for tasks like intent
detection and sentiment analysis to automate responses and improve customer
satisfaction. Fine-tuning models on customer service dialogues will enhance
their accuracy in handling queries.
● Financial Analysis: Fine-tune RoBERTa or FinBERT for analyzing financial
documents, news, or reports. These models can help with sentiment analysis,
financial trend prediction, and event extraction.
● Healthcare Support: Adapt BioBERT or ClinicalBERT for processing
medical records, diagnostic reports, and research papers. These models can be
fine-tuned to understand medical terminology and provide context-specific

Certainly! Here are the answers to the remaining questions:

19. Training Large Language Models: Primary Obstacles and Solutions

Training large language models (LLMs) requires significant resources, with primary
obstacles being:

● Data Requirements: LLMs require vast amounts of diverse, high-quality data

to achieve generalization. Collecting such data, especially domain-specific
data, can be time-consuming and expensive.
● Computational Power: Training LLMs demands considerable computational
resources, often requiring specialized hardware like GPUs or TPUs for months
at a time, making it costly.
● Environmental Impact: The energy consumption of large-scale model training
can have a substantial environmental impact.

Solutions:

● Data Augmentation and Synthetic Data: These techniques help overcome

data scarcity, especially for niche domains or low-resource languages.
● Model Optimization: Techniques like model distillation (reducing model size
without losing much performance) and parameter sharing can help reduce the
computational burden.
● Efficient Training Methods: Transfer learning allows for models to be
fine-tuned rather than trained from scratch, reducing both time and
computational costs.

20. Guidelines to Minimize Risks in Model Deployment

Organizations deploying language models must implement specific guidelines to

minimize risks like bias, data privacy violations, and security threats:
 ● Data Privacy and Security: Follow privacy regulations (such as GDPR) and
ensure data is anonymized to protect user confidentiality. Implement strict data
access controls and encryption.
● Bias Audits: Regularly audit the models for bias, especially when deployed in
sensitive fields (e.g., healthcare or criminal justice). Utilize bias detection
frameworks and ensure fairness by retraining models with diverse data.
● Transparency and Accountability: Clearly document model development
processes, training data, and assumptions. Develop mechanisms for
accountability in decision-making, such as model explainability features.
● Ethical Standards: Follow ethical guidelines and ensure models do not harm
individuals or communities. This includes ensuring the safety of users and
preventing model misuse.

These measures can be applied consistently across sectors to ensure safer and more
responsible model deployment.

21. Bias in AI: Origins and Examples in Pre-Trained Models

Bias in AI models can emerge from several sources:

● Bias in Training Data: If the training data contains biased or imbalanced

representations (e.g., underrepresentation of certain demographics), the model
can learn and perpetuate these biases.
● Historical Bias: Many datasets reflect historical inequalities (e.g., gender,
race), which the model inadvertently learns to replicate.
● Sampling Bias: Models trained on data from one group or region may fail to
generalize to others, leading to biased outputs.

Example: A Hugging Face model trained on news data may reflect biased views
toward certain political ideologies, with a skewed portrayal of events or individuals
based on the news sources used for training.

Consequences: Bias in models can harm marginalized groups by reinforcing

stereotypes or making discriminatory decisions.

22. Bias in NLP Tasks: Sentiment Analysis, Language Translation, and Text
Generation

Bias in NLP tasks, such as sentiment analysis, language translation, and text
generation, can lead to undesirable outcomes:

● Sentiment Analysis: Bias in sentiment analysis can result in the

misinterpretation of emotions or sentiments in text from different social or
cultural contexts. For example, reviews written by users from minority groups
may be unfairly classified as negative, even when the sentiment is neutral or
 positive.
● Language Translation: Bias in machine translation systems may result in the
translation of gender-neutral terms into gendered language, potentially
perpetuating stereotypes (e.g., translating "doctor" as "he" instead of "they").
● Text Generation: Text generation models may output harmful or biased
content if not properly regulated. For instance, a model generating job
descriptions may inadvertently reflect gender biases, preferring male candidates
over female candidates.

Risks in Real-World Applications:

● In hiring processes, biased sentiment analysis or text generation may lead to

unfair job screening.
● In legal assessments, biased translations or analysis could lead to unfair
judgments or discriminatory legal outcomes.
● In customer service, biased language models could provide harmful or
inappropriate responses, damaging customer relationships.

23. Ethical Concerns in Critical Domains: Healthcare and Criminal Justice

Using biased AI models in fields like healthcare and criminal justice can have
profound ethical consequences:

● Healthcare: If AI models trained on biased datasets make medical diagnoses or

treatment recommendations, it could lead to harmful outcomes for
underrepresented groups. For instance, a model might underperform when
diagnosing diseases in women or people of color because it has been trained
primarily on data from men or white individuals.
● Criminal Justice: AI models used for risk assessments in the criminal justice
system might reinforce biases present in the training data, potentially leading to
unfair sentencing or parole decisions, especially for minority groups.

Impact on Trust: If AI systems are perceived as unfair or discriminatory, it may lead

to a lack of trust in these technologies, especially in critical domains. This could
hinder the adoption of AI and affect public confidence in systems that rely on AI
decision-making.

24. Risks of Bias in Healthcare and Judicial Assessments

Bias in AI models used for healthcare or judicial assessments can lead to significant
risks for both individuals and society:

● Healthcare: Inaccurate medical diagnoses or treatment recommendations due

to bias could affect patient outcomes, particularly for those in underrepresented
or marginalized communities. For instance, biased diagnostic tools might
 misdiagnose women or people of color because of the lack of diverse data used
during training.
● Judicial Assessments: AI used for assessing recidivism risk or sentencing
decisions could result in unfair treatment based on race, gender, or
socioeconomic status, disproportionately affecting minority groups.

Risks:

● Individual Impact: Unjust decisions in healthcare or legal systems can lead to

physical harm, wrongful convictions, or unnecessary incarcerations.
● Societal Impact: The amplification of social inequalities can deepen existing
societal disparities and erode trust in AI systems used in these domains.

25. Mitigating Bias in AI Models for Healthcare and Criminal Justice

To mitigate bias in AI models, especially in high-stakes domains such as healthcare

and criminal justice, several strategies should be applied:

● Data Audits: Regular audits of the training data to identify and correct for
biases in the datasets.
● Debiasing Algorithms: Implement algorithms designed to remove or reduce
bias, ensuring fairer decision-making.
● Diverse and Representative Data: Ensure that training datasets are diverse,
inclusive, and represent a wide range of demographics, experiences, and
perspectives.
● Rigorous Testing: Test the models thoroughly in real-world scenarios to
evaluate their fairness, performance, and accuracy across diverse groups.

Example: In healthcare, using balanced datasets that include data from different
ethnic groups ensures that the model performs well across all demographic groups,
avoiding discrimination.

26. Hugging Face Transformers Architecture Overview

Hugging Face Transformers is a widely used library that simplifies the deployment of
NLP models for various tasks:

● Core Components:
○ Pre-trained Models: The library provides access to state-of-the-art
transformer models (like BERT, GPT, RoBERTa, T5, etc.) for a variety
of tasks, including text classification, translation, and summarization.
○ Tokenizers: Tokenizers convert text into a format that models can
process, breaking down text into tokens (words, subwords, etc.).
○ Pipelines: Hugging Face offers a simple API for running tasks like text
generation, classification, and translation. These pipelines abstract away
 the complexities of preprocessing and postprocessing.
● Flexibility: The library allows users to fine-tune pre-trained models on custom
datasets, making it adaptable for a wide range of NLP applications.

27. Model Selection, Fine-Tuning, and Hyperparameter Tuning in Hugging Face

When fine-tuning a model using Hugging Face:

1. Load a Pre-Trained Model: Choose a suitable pre-trained model (e.g., BERT

for text classification) and load it from the Hugging Face Model Hub.
2. Prepare Data: Tokenize the dataset (converting text into input tokens) and
split it into training and validation sets.
3. Fine-Tuning: Adjust the model’s weights using a supervised learning
approach, where the model learns from the labeled training data. Fine-tuning
can be done using methods like gradient descent.
4. Hyperparameter Tuning: Adjust hyperparameters like learning rate, batch
size, and the number of training epochs to improve model performance.
5. Evaluate: After training, evaluate the model on a validation dataset to check
accuracy, precision, recall, and other relevant metrics.

28. Speech Recognition and Machine Translation Models from Hugging Face

1. Speech Recognition Models:

○ Wav2Vec 2.0: A state-of-the-art model for automatic speech recognition.
It can transcribe spoken language into text with high accuracy.
○ DeepSpeech: An open-source speech-to-text model that can be easily
integrated with Hugging Face.
2. Machine Translation Models:
○ MarianMT: A versatile translation model supporting many language
pairs, ideal for multilingual use.
○ MBart: Another powerful model for machine translation, particularly
effective for translation between multiple languages and diverse
language pairs.

1. What are the three fine-tuning methods you implemented, and why are they
important?

Answer: The three fine-tuning methods are:

● Full Fine-Tuning: This involves adjusting all the parameters of the model
based on the new dataset. It provides the most flexibility but requires a large
amount of data and computing resources.
● Parameter-Efficient Fine-Tuning (PEFT): This focuses on adjusting only a
subset of the model's parameters, allowing faster and more efficient training
with less data while preserving the model's core abilities.
 ● Few-Shot/Prompt-Based Tuning: This method involves training the model
with only a few examples and relies on prompt engineering to guide the
model’s predictions. It is highly efficient when large datasets aren't available.

These methods are important as they allow customization of models to meet specific
departmental needs with varying resource constraints.

2. How did you fine-tune the model for the Customer Service department?

Answer: For the Customer Service department, the model was fine-tuned to handle
sentiment analysis, question answering, and chat summarization. The fine-tuning
process incorporated datasets that included informal language, slang, and customer
support-specific terminology. Full fine-tuning was used initially to adjust the entire
model to handle the variety of language encountered in customer interactions.

3. What challenges did you face when fine-tuning the model for Financial Analysis?

Answer: One challenge in fine-tuning for Financial Analysis was ensuring that the
model could handle industry-specific terminology and unstructured financial data
while remaining accurate over time. PEFT was a useful method here, as it allowed
fine-tuning on financial datasets without completely overhauling the base model,
minimizing the risk of overfitting and concept drift in a domain where data evolves
continuously.

4. How did you ensure the Healthcare Support model adhered to privacy and accuracy
concerns?

Answer: The Healthcare Support model required careful attention to data privacy and
avoiding overfitting due to limited datasets. We applied Few-Shot/Prompt-Based
Tuning to work with small, privacy-conscious healthcare datasets. This method
allowed the model to recognize medical entities and extract diagnosis-related
information effectively while avoiding overfitting by training with fewer examples.

5. What datasets did you use for each department's fine-tuning, and how did they
impact the results?

Answer:

● Customer Service: Datasets included customer chat logs, informal language

corpora, and sentiment-labeled conversations. These helped the model learn
how to manage various customer support scenarios and recognize sentiment in
diverse language styles.
● Financial Analysis: Financial reports, stock market data, and industry-specific
documents were used to fine-tune the model. This enabled the model to handle
trend prediction and financial summary generation accurately.
● Healthcare Support: Medical case reports, de-identified patient data, and
healthcare terminology corpora were used. Since healthcare data is sensitive,
few-shot learning was key in ensuring that the model handled medical terms
while respecting privacy.
6. How did the fine-tuning methods compare in terms of performance and efficiency
across departments?

Answer:

● Full Fine-Tuning provided the highest accuracy across all departments but
required more resources and time, making it less efficient for smaller datasets.
● Parameter-Efficient Fine-Tuning (PEFT) balanced performance and
efficiency, especially for Financial Analysis, where maintaining the model’s
general knowledge while adapting to domain-specific terms was critical.
● Few-Shot/Prompt-Based Tuning excelled in Healthcare Support, where data
privacy and the small size of available datasets made it the most practical
choice. It performed well with minimal data but required careful prompt design
to guide the model.

7. How did you ensure that each department's model remained generalizable while
being fine-tuned for specific tasks?

Answer: We applied techniques such as early stopping and regularization during

fine-tuning to prevent overfitting to the new data. For PEFT and Few-Shot Tuning,
we limited the changes to only a subset of the model parameters, which helped retain
the model's general knowledge while adapting to specific departmental tasks.

8. What role did prompt engineering play in Few-Shot/Prompt-Based Tuning, and how
was it applied in this project?

Answer: Prompt engineering was crucial for guiding the model's outputs in Few-Shot
Tuning, particularly for Healthcare Support. By designing prompts that clearly
defined the task (e.g., extracting medical entities or summarizing patient diagnoses),
the model was able to perform well even with limited examples. This method allowed
for flexibility in adapting to different tasks with minimal data.

These answers should help explain the fine-tuning methods and challenges associated
with each department’s requirements during your project viva.