NeuroCast — Global NLP Intelligence Platform

NeuroLogic '26 · Global NLP Datathon · GGITS AIML Dept · April 25, 2026

Overview

NeuroCast is a full-stack NLP platform solving all three NeuroLogic '26 challenges:

• Challenge 1: Disaster Tweet Classification (Metric: Macro F1-Score)
• Challenge 2: Fake News Detection (Metric: Accuracy)
• Challenge 3: Multilingual Toxic Comment Classification (Metric: ROC-AUC)

Results

Challenge	Model	Metric	Score
1 — Disaster Tweets	TF-IDF + Logistic Regression	Macro F1	95]
2 — Fake News	TF-IDF + Logistic Regression	Accuracy	98
3 — Toxic Comments	TF-IDF + Logistic Regression	ROC-AUC	97

Approach

Challenge 1 — Disaster Tweet Classification

• Dataset: CrisisLexT26 (25,933 tweets, 2012–2013)
• Labels: 1 = Informative, 0 = Not Informative
• Model: TF-IDF (max_features=15000, ngram_range=(1,2), sublinear_tf=True) → Logistic Regression (C=1.0, class_weight='balanced')
• Train/Test Split: 80/20 stratified
• Metric: Macro F1-Score (handles class imbalance: 16019 vs 9914)

Challenge 2 — Fake News Detection

• Dataset: Fake and Real News Dataset (23,893 articles)
• Labels: "TRUE" = Real, "FALSE" = Fake
• Feature Engineering: title + text concatenated before vectorization
• Model: TF-IDF (max_features=20000, ngram_range=(1,2)) → Logistic Regression (C=5.0)
• Train/Test Split: 80/20 stratified
• Metric: Overall Accuracy

Challenge 3 — Multilingual Toxic Comments

• Dataset: HuggingFace Multilingual Toxicity Dataset (9,000 samples, EN + HI)
• Labels: 0 = Non-toxic, 1 = Toxic
• Model: TF-IDF (max_features=15000, sublinear_tf=True) → Logistic Regression (C=2.0, class_weight='balanced')
• Train/Test Split: 80/20 stratified
• Metric: Mean ROC-AUC

How to Reproduce Results

Prerequisites

• Python 3.11+
• Node.js 18+
• pip

Setup & Run

Backend:

cd backend
pip install -r requirements.txt
python train_models.py    # trains all 3 models, saves .pkl files
python app.py             # starts API on port 5000

Frontend:

cd frontend
npm install
npm run dev               # starts UI on port 3000

Generate Submission CSVs:

1. Visit http://localhost:3000/dashboard
2. Click "Download All Predictions (ZIP)"
3. Unzip — contains disaster_predictions.csv, fakenews_predictions.csv, toxic_predictions.csv

Dataset Usage

Place the following files in backend/data/ before running train_models.py:

• Disaster_with_label.csv + Disaster_no_label.csv
• fakenews_with_labels.csv + FakeNews_no_labels.csv
• toxic_labeled.xlsx + toxic_no_label_evaluation.xlsx

(Available from the NeuroLogic '26 official Google Drive)

Tech Stack

Layer	Technology
Backend API	Flask 3.0, Flask-CORS
ML	scikit-learn, TF-IDF, Logistic Regression
Data	pandas, numpy, openpyxl, joblib
Frontend	React 18, Vite, Tailwind CSS
Charts	Chart.js, react-chartjs-2
Animations	Framer Motion
Language Detection	langdetect

Prediction Files

• backend/predictions/disaster_predictions.csv — labels: 0 or 1 (integer)
• backend/predictions/fakenews_predictions.csv — labels: "TRUE" or "FALSE" (string)
• backend/predictions/toxic_predictions.csv — labels: 0 or 1 (integer)