7043SCN Generative AI and Reinforcement Learning Assignment 1 – Comparative Fine-Tuning of Phi-3 Mini for E-commerce Product Question Answering

Student: J. Panfilova
Module: 7043SCN Generative AI & Reinforcement Learning
Academic year: 2026/27
Task: Task 1 – LLM Fine-Tuning
Status: Milestone 1 – Project setup and baseline preparation

1. Project Summary

This project investigates whether parameter-efficient fine-tuning can improve the ability of an open-source Large Language Model (LLM) to answer factual questions about e-commerce product listings.

The selected application is eBay product question answering. The model is given contextual information from an eBay product listing and is asked a factual question about one of the product attributes.

Examples include:

Is this product discontinued?
Who is the manufacturer?
What is the model name of this product?
What is the listed price?
How many reviews does the seller have?
What is the seller rating?
Is the product currently in stock?
Is this product listed as used?

Research Question

How effectively can QLoRA fine-tuning improve Phi-3 Mini's ability to answer factual questions about eBay product attributes compared with the pretrained model?

Hypothesis

QLoRA fine-tuning will improve the factual accuracy and relevance of Phi-3 Mini's answers to questions about eBay product attributes compared with the untuned pretrained model.

The main comparison will therefore be between:

the pretrained Phi-3 Mini model used as the baseline; and
the same model adapted using QLoRA.

Full fine-tuning is considered technically infeasible on the available GPU memory, so QLoRA is being investigated as the parameter-efficient adaptation method.

2. MODELING

The selected Hugging Face model is:

microsoft/Phi-3-mini-4k-instruct

Phi-3 Mini is a decoder-only instruction-following language model and is therefore suitable for the proposed question-answering formulation.

The model has approximately 3.8 billion parameters in its original model configuration and supports a 4K-token context window.

The model is loaded using 4-bit quantization to make experimentation possible on the available GPU.

The planned QLoRA configuration uses:

4-bit NF4 quantization
double quantization
LoRA adapters
rank r = 8
LoRA scaling factor alpha = 16
LoRA dropout 0.05
all linear layers as LoRA targets

During preliminary testing, FP16 computation produced numerical instability with NaN logits. FP32 computation was therefore selected for the QLoRA experiments because it produced finite and stable outputs.

3. Dataset

The project uses the public Hugging Face dataset:

montassar-souli/ebay_productlisting

The original dataset contains approximately 30,000 eBay product listings and 26 columns.

During data preparation:

completely empty columns were removed;
broken-link entries were removed;
product-level splitting was used to prevent the same product appearing across different splits;
a deterministic random seed of 42 was used.

The resulting cleaned dataset contains 29,552 product listings:

Split	Products
Train	23,641
Validation	2,955
Test	2,956
Total	29,552

The three splits have no overlapping product IDs.

3.1 Product Attributes

The original eBay listings contain attributes such as:

Title
Model Number
UPC
Manufacturer
Model Name
Price
Stock
Carrier
Color Category
Internal Memory
Screen Size
Discontinued
Seller Rating
Seller Number of Reviews
Used

Some attributes are sparsely populated in the original data. Missing values are therefore handled as missing context rather than being replaced with artificial values.

3.2 Contextual Question Answering

Eight factual question-answer pairs are generated for each product.

The target fields are:

Discontinued
Manufacturer
Model Name
Price
Seller Num Of Reviews
Seller Rating
Stock
Used

The question-answer data is converted into a contextual product question-answering format.

For each example, the model receives the available product listing attributes except for the attribute being directly asked about.

For example, for a manufacturer question:

Product listing:
Title: ...
Model Name: ...
Price: ...
Stock: ...
Discontinued: ...
Seller Rating: ...

Question:
Who is the manufacturer of this product?

The expected manufacturer is kept separately as the target answer and is not included as a separate Reference: field in the model input.

This formulation allows the model to use the surrounding product information while requiring it to generate the answer to the specific factual question.

Missing source attributes are omitted from the contextual prompt. When the original dataset does not specify an answer, the target is represented as Not specified.

The complete contextual training dataset is stored as:

qa/train_qa_contextual_v2.csv

3.3 QLoRA Training Subset

The complete training split contains 189,128 QA examples (23,641 products × 8 questions).

Because of the available GPU resources and the computational cost of training the complete QA dataset, a reproducible product-level subset is used for the initial QLoRA experiment.

The subset contains:

1,250 products
8 questions per product
10,000 QA examples
1,250 examples for each target field
random seed 42

The corrected contextual training subset is stored as:

qa/qlora_train_subset_contextual_10000.csv

The subset is selected at product level so that all eight questions belonging to a selected product remain together.

4. Evaluation Plan

A held-out evaluation set will be constructed from the test split.

The same evaluation examples will be presented to both:

the pretrained Phi-3 Mini baseline; and
the QLoRA fine-tuned model.

This provides a direct comparison between the untuned and adapted models.

The evaluation will use task-specific factual accuracy together with appropriate text-generation metrics where useful.

Quantitative evaluation will consider measures such as:

exact or normalized factual accuracy;
field-specific accuracy;
token-level similarity;
ROUGE-L where appropriate.

Qualitative analysis will also examine representative correct and incorrect answers from the two models.

The evaluation prompts will contain the same contextual product information for both models so that the comparison measures the effect of fine-tuning rather than a difference in input information.

No final performance claim is made at this milestone.

5. Environment and Hardware

Experiments are conducted using Google Colab.

Current tested environment:

GPU: NVIDIA Tesla T4
GPU memory: 14.56 GB
CUDA: 13.0
PyTorch: 2.11.0+cu130
Transformers: 5.17.0
PEFT: 0.21.0
bitsandbytes: 0.50.2
TRL: 1.14.1
Datasets: 4.8.5

The available GPU memory is insufficient for practical full fine-tuning of the approximately 3.8B-parameter Phi-3 Mini model with optimizer states and activations.

Therefore, the project uses QLoRA as the efficient fine-tuning approach and compares the resulting model against the pretrained baseline.

The 4-bit NF4 configuration was successfully tested on the T4 GPU.

6. Software and Reproducibility

The project is implemented in Python.

The main libraries include:

PyTorch
Hugging Face Transformers
Hugging Face Datasets
PEFT
TRL
bitsandbytes
Accelerate
scikit-learn
pandas
NumPy

A deterministic random seed of 42 is used for the main dataset splitting and sampling procedures.

The project directory contains the notebook, data-processing code, QA data, evaluation outputs, and model-related files.

Large datasets, model weights, checkpoints, and generated outputs are excluded from Git where appropriate.

7. Current Milestone 1 Status

At Milestone 1, the following project components have been established:

open-source model selected: microsoft/Phi-3-mini-4k-instruct;
e-commerce product QA application defined;
public eBay product dataset selected;
cleaned product-level train/validation/test splits created;
eight factual QA fields defined;
contextual QA formulation implemented;
reproducible 10,000-example QLoRA training subset created;
available GPU environment tested;
4-bit NF4 model loading verified;
QLoRA identified as the primary efficient fine-tuning method;
baseline and fine-tuned model comparison planned using the same held-out contextual evaluation set.

The project is continuing toward the corrected baseline evaluation and QLoRA training experiment.

8. Project Structure

7043SCN_Task1_eBay/
├── data/
├── qa/
├── results/
│   └── figures/
├── models/
│   ├── baseline/
│   └── qlora/
├── src/
│   ├── __init__.py
│   └── load_data.py
├── examples/
│   └── sample_products.csv
├── notebooks/
│   └── 7043SCN_Task1_eBay.ipynb
├── .gitignore
├── README.md
└── requirements.txt

9. Planned Experimental Workflow

Public eBay product dataset
          ↓
Data cleaning and product-level split
          ↓
Factual QA generation
          ↓
Contextual product QA formulation
          ↓
Pretrained Phi-3 Mini baseline
          ↓
QLoRA fine-tuning
          ↓
Same held-out evaluation questions
          ↓
Quantitative + qualitative comparison

The final analysis will examine whether QLoRA produces measurable improvements in factual product-attribute question answering compared with the pretrained model.

panfilovaj/milestone2_commit

About

Resources

Uh oh!

Stars

Watchers

Forks

Releases

Contributors 1

Languages