FLAN-T5 LoRA — Lay Radiology Summarization (BioLaySumm Subtask 2.1) - 47451773

[TPGCIG]

Author: Tristan Green (s4745177)
Target branch: topic-recognition
Project folder: recognition/Project13-TristanGreen

Summary

This PR contributes a parameter-efficient fine-tuning of FLAN-T5-base with LoRA for layperson summarization of radiology reports. It includes:

• Reproducible training (train.py), evaluation and generation (predict.py and chat.py)

• Modular components (modules.py) and documented configuration.

• Plots + metrics (loss curve, validation ROUGE trajectory, test bar chart).

• A comprehensive README.md.

The algorithm trains on the provided training set, validates on a held-out split for model selection, and reports ROUGE-1/2/L/Lsum on a held-out test set.

Recognition problem & difficulty

Task: Expert→lay summarisation (long-form seq2seq, automatic evaluation via ROUGE)

Setup and Run Script

Setup and running is highlighted explicitly in the README.md, however, the script to run the code with default parameters is:

python train.py

Thank you,
Tristan (s4745177)

[Claire1217]

Recognition Problem : total : 17
Solves problem: Overall good work. Some problems listed here: Files in 5) outputs doesn't align with what you acturally have in the repo. Why use a chat interface? What's the difference between chat and inference? Why you use LoRA? How much space used? Label masking to -100 is what's usually done in gpt2 tokenizer not in T5 tokenizer. "r=8 is a common sweet spot for T5‑base" How do you know? Have you tried different r? (3)
Implementation functions : The code seems to be functional(2)
Good design: good (1)
Commenting: Well commented (1)
Difficulty: Hard (10)

[gayanku]

Marking

Good/OK/Fair Practice (Design/Commenting, TF/Torch Usage)
	Good design and implementation.
	Spacing and comments.
	No Header blocks.	-1
Recognition Problem
	OK solution to problem.	-1
	Driver Script present.
	File structure present.
	Good Usage & Demo & Visualisation & Data usage.
	Module present.
	Commenting present.
	No Data leakage found.
	Difficulty : Hard. Hard Difficulty : LLM
Commit Log
	Good Meaningful commit messages.
	Good Progressive commits.
Documentation
	Readme :Acceptable.	-2
	Model/technical explanation :Good.
	Description and Comments :Good.
	Markdown used and PDF submitted.
Pull Request
	Successful Pull Request (Working Algorithm Delivered on Time in Correct Branch).
	No Feedback required.
	Request Description is good.
TOTAL		-4

Marked as per the due date and changes after which aren't necessarily allowed to contribute to grade for fairness.
Subject to approval from Shakes

[TPGCIG]

I just want to add that I had an approved extension, I emailed this to Shakes. All elements of the PR were submitted in time of the extension.