StudyLens

AI-powered lecture review platform that consolidates slides, transcripts, and notes into structured summaries with RAG-based Q&A.

Built for AIPI 540 NLP Module Project, Duke University

• Live App: https://studylens.up.railway.app/
• Repository: https://github.com/SophiaYifei/StudyLens.git
• Team: Yifei Guo (SophiaYifei), Sharmil Nanjappa (SharmilNK)

---

Project Overview

StudyLens takes three types of input per lecture — slide decks (PPTX), auto-generated transcripts, and student notes — and produces structured summaries for exam review. We evaluate 9 model configurations across three paradigms (naive baselines, classical ML, deep learning) on 10 lecture topics from two Duke graduate courses (AIPI 540: Deep Learning, AIPI 520: Machine Learning).

The deployed web application features two modes:

• Review Mode: Real-time AI summarization via DeepSeek API, pre-computed fine-tuned Qwen summaries for comparison, and a RAG-based Q&A chatbot grounded in lecture materials
• Lecture Mode: PDF slide viewer with integrated note-taking

---

Repository Structure

StudyLens/
├── README.md                        <- This file
├── requirements.txt                 <- Base environment (BART, Long-T5, evaluation)
├── requirements-llm.txt             <- LLM environment (Qwen2.5-7B)
├── requirements-app.txt             <- Web app dependencies
├── .env.example                     <- API key template (copy to .env)
├── .gitignore
├── LICENSE
├── Makefile
├── main.py                          <- Data processing pipeline entry point
├── app.py                           <- FastAPI web application backend
├── Procfile                         <- Railway deployment config
├── railway.toml                     <- Railway build settings
├── static/
│   └── index.html                   <- Frontend (single-file HTML/CSS/JS)
├── scripts/
│   ├── model.py                     <- All summarization models (naive, TF-IDF, neural, LLM)
│   ├── make_dataset.py              <- Data loading and transcript denoising
│   ├── build_features.py            <- Chunking, embedding, topic assignment
│   ├── naive.py                     <- First-5-slides heuristic baseline
│   ├── eval.py                      <- Unified evaluation (ROUGE-L, BERTScore, NLI)
│   ├── finetune_qwen.py             <- QLoRA fine-tuning pipeline
│   └── plot_from_csv.py             <- Generate all report figures from eval CSV
├── models/
│   └── qwen7b-lora/                 <- Fine-tuned LoRA adapter weights (~40MB)
├── data/                            <- (gitignored, downloaded from HuggingFace at runtime)
│   ├── raw/                         <- Slides, transcripts, notes, files are removed due to privacy
│   ├── processed/                   <- Combined _ori.txt files, files are removed due to privacy
│   ├── reference/                   <- GPT-5.2-generated reference summaries
│   ├── outputs/                     <- Model-generated summaries
│   └── eval/                        <- Evaluation results and plots
└── notebooks/                       <- Exploration notebooks (not graded)

---

Models

All models are implemented in scripts/model.py (sharing a common BaseSummarizer interface) except the first-5-slides heuristic which is in scripts/naive.py.

Requirement	Class / File	--model flag	Description
Naive baseline	FirstSentenceSummarizer	first5	First sentence from first 5 slides
Naive baseline	RandomExtractiveSummarizer	random	Randomly select 15 sentences
Classical ML	TFIDFExtractiveSummarizer	tfidf	TF-IDF sentence ranking, top 15
Deep learning	BARTSummarizer	bart	BART-CNN, 1024 tokens, news-trained
Deep learning	LongT5Summarizer	longt5	Long-T5, 16384 tokens, book-trained
Deep learning	BARTSamsumSummarizer	bart-samsum	BART-SAMSum, 1024 tokens, dialogue-trained
Deep learning	LEDArxivSummarizer	led-arxiv	LED-arXiv, 16384 tokens, academic paper-trained
Deep learning	QwenSummarizer	qwen7b	Qwen2.5-7B, 128K context, 4-bit quantized
Deep learning	finetune_qwen.py	via script	QLoRA fine-tuned Qwen on lecture data

---

Environment Setup

This project uses two environments due to dependency conflicts between encoder-decoder models and Qwen.

Base environment (BART, Long-T5, BART-SAMSum, LED-arXiv, Evaluation)

pip install -r requirements.txt

LLM environment (Qwen2.5-7B, requires NVIDIA GPU with 16GB+ VRAM)

pip install -r requirements-llm.txt

Why two environments? BART/Long-T5 require transformers==4.41.2, while Qwen2.5-7B requires transformers>=4.45.0. Running both in one environment causes version conflicts.

---

Running the Pipeline

Step 1: Process raw data

python main.py

This loads raw files from data/raw/, denoises transcripts, and produces 10 combined _ori.txt files in data/processed/.

Step 2: Generate summaries

# Base environment:
python scripts/model.py --model first5
python scripts/model.py --model random
python scripts/model.py --model tfidf
python scripts/model.py --model bart --strategy both
python scripts/model.py --model longt5 --strategy both
python scripts/model.py --model bart-samsum --strategy both
python scripts/model.py --model led-arxiv --strategy both

# LLM environment (requires GPU):
python scripts/model.py --model qwen7b
python scripts/model.py --model qwen7b --output_ratio 0.01
python scripts/model.py --model qwen7b --output_ratio 0.03

Step 3: Fine-tune Qwen (LLM environment, A100 GPU)

python scripts/finetune_qwen.py --mode prepare
python scripts/finetune_qwen.py --mode train --epochs 5
python scripts/finetune_qwen.py --mode inference

Step 4: Evaluate all models

python scripts/eval.py              # ROUGE-L + BERTScore
python scripts/eval.py --run-nli    # also compute NLI (slower)

Results saved to data/eval/evaluation_results_all.csv and data/eval/evaluation_averages.csv.

Step 5: Generate plots

python scripts/plot_from_csv.py

Figures saved to data/eval/plots/.

---

Running on Google Colab

Due to GPU requirements, we recommend Colab with an A100 GPU. Use two separate notebooks to avoid dependency conflicts.

Mount Drive (both notebooks):

from google.colab import drive
drive.mount('/content/drive')
import os
os.chdir("/content/drive/{your directory}")

Notebook A (base environment):

!pip install -r requirements.txt
!python scripts/model.py --model bart --strategy both
!python scripts/model.py --model longt5 --strategy both
!python scripts/model.py --model bart-samsum --strategy both
!python scripts/model.py --model led-arxiv --strategy both
!python scripts/eval.py --run-nli

Notebook B (LLM environment, separate session):

!pip install -r requirements-llm.txt
!python scripts/model.py --model qwen7b
!python scripts/finetune_qwen.py --mode train --epochs 5
!python scripts/finetune_qwen.py --mode inference

---

Web Application

The deployed app is at https://studylens.up.railway.app/

Local development

pip install -r requirements-app.txt
export OPENROUTER_API_KEY=your-key
python app.py

Deployment (Railway)

The app auto-downloads data from a private HuggingFace dataset repo (YifeiGuo/studylens-data) on first startup. Required environment variables on Railway:

• OPENROUTER_API_KEY — for real-time summarization and chat
• HF_TOKEN — for downloading data from private HuggingFace repo

---

API Keys

Copy the example environment file and add your credentials:

cp .env.example .env

• ANTHROPIC_API_KEY — for ClaudeSummarizer (optional; all other models work without it)
• OPENROUTER_API_KEY — for the web app's real-time AI features

The .env file is in .gitignore and will never be committed.

---

AI Attribution

Code co-authored with Claude (Anthropic, https://claude.ai) for structural design, debugging, and documentation. All files contain attribution headers.