Translating Expert Radiology Reports into Layman Summaries using pre-trained Flan-T5 Model. Task #13

[piradiusquared]

Expert to Layman Radiology Reports

Overview:

This project aims to fine-tune an existing encoder-decoder LLM - the Flan-T5 in specifically summarising data from the BioLaySumm 2025 dataset.
Fine-tuning was performed on the rangpur cluster, with full training on the NVDIA A100 GPU.

Files:

constants.py - File containing all parameters and other constant variables
dataset.py - Custom dataset loader using Pandas for model training and evaluation
modules.py - Importer for the pre-trained Flan-T5 model
predict.py - Model benchmarking on unseen validation data split
requirements.txt - List of all dependencies and their respective version
train.py - Custom training and evaluation loop on the train data split
assets/ - Folder for all README required images
runners/ - Folder containing train and benchmark scripts for use on rangpur

Result of Model Fine-tuning:

rouge scores:

• rouge1: 71.19
• rouge2: 51.83
• rougeL: 65.80
• rougeLsum: 65.82

perplexity scores (extra metrics):

• On average, the fine-tuned model is 2 - 4 times more confident in selecting the next appropriate token (word) for summaries.
• Average perplexity score is consistently < 3

[Claire1217]

Recognition Problem : total : 20
Solves problem: Good data augmentation design. Very good reasoning from test run 1 to test run 2. (5)
Implementation functions : Well-organized codes. Good practice to have your constants in an individual file(3)
Good design: good (1)
Commenting: Well commented (1)
Difficulty: Hard (10)

[wangzhaomxy]

s4885380

[gayanku]

Marking

Good/OK/Fair Practice (Design/Commenting, TF/Torch Usage)
	Good design and implementation.
	Spacing and comments.
	No Header blocks.	-1
Recognition Problem
	Good solution to problem.
	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 :Good.
	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		-1

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