NeuroLogic '26 Datathon — NLP Solutions

Author: Kishan Kumar GitHub: https://github.com/techwithkishan/Neuro-Logic26-Datathon Event: NeuroLogic '26 Datathon, GGITS — April 25, 2026

---

🏆 Results Summary

Challenge	Task	Metric	Score
Challenge 1	Disaster Tweet Classification	Macro F1	0.8955
Challenge 2	Fake News Detection	Accuracy	99.87%
Challenge 3	Multilingual Toxic Comments	ROC-AUC	0.9922

---

🚨 Challenge 1 — Disaster Tweet Classification

Model: vinai/bertweet-base — pretrained on 850M tweets Dataset: 25,933 train / 2,000 test Labels: Informative / Not Informative

Approach:

• Combined Tweet Text + Information Source + Information Type
• Fine-tuned BERTweet for 3 epochs
• 90/10 train-validation split, seed=42
• max_length=128, batch_size=32, fp16=True

Validation Results:

Class	Precision	Recall	F1
Informative	0.92	0.93	0.92
Not Informative	0.87	0.86	0.87
Macro F1			0.8955

BERTweet was specifically chosen due to its pretraining on large-scale Twitter data, making it highly effective for domain-specific tasks like disaster tweet classification.

---

📰 Challenge 2 — Fake News Detection

Model: roberta-base Dataset: 18,176 train / 1,541 test Labels: True / False

Approach:

• Combined title + [SEP] + text + [SEP] + subject
• Fine-tuned RoBERTa for 3 epochs
• 90/10 train-validation split, seed=42
• max_length=256, batch_size=16, fp16=True

Validation Results:

Class	Precision	Recall	F1
False	1.00	1.00	1.00
True	1.00	1.00	1.00
Accuracy			99.87%

---

🌐 Challenge 3 — Multilingual Toxic Comment Classification

Model: xlm-roberta-base — supports 100 languages Dataset: 9,000 train / 1,000 test Labels: 0 (Non-Toxic) / 1 (Toxic) Languages: Hindi + English

Approach:

• Fine-tuned XLM-RoBERTa for 3 epochs
• 90/10 train-validation split, seed=42
• max_length=128, batch_size=16, fp16=True
• ROC-AUC computed using softmax probabilities

Validation Results:

Class	Precision	Recall	F1
Non-Toxic (0)	0.97	0.96	0.96
Toxic (1)	0.96	0.97	0.96
ROC-AUC			0.9922

---

▶️ How to Reproduce

1. Open Google Colab — enable T4 GPU (Runtime → Change runtime type → T4 GPU)
2. Upload dataset files to Colab
3. Install dependencies: pip install -r requirements.txt
4. Run notebooks top to bottom:
   • challenge1/C1_Disaster_Tweets.ipynb
   • challenge2/C2_Fake_News.ipynb
   • challenge3/C3_Toxic_Comments.ipynb
5. Prediction files will be saved automatically

All random seeds set to 42 for reproducibility

---

📁 Repository Structure

• 📁 challenge1/
  • NL'26_Datathon_Challenge_1.ipynb
  • Disaster_no_label.csv
• 📁 challenge2/
  • NL'26_Datathon_Challenge_2.ipynb
  • FakeNews_no_labels.csv
• 📁 challenge3/
  • NL'26_Datathon_Challenge_3.ipynb
  • toxic_no_label_evaluation.xlsx
• 📁 screenshots/
  • c1_results.png
  • c2_results.png
  • c3_results.png
• README.md
• requirements.txt

---

🛠️ Technology Stack

• Python 3.12
• HuggingFace Transformers
• PyTorch
• scikit-learn
• Google Colab T4 GPU
• pandas, numpy, openpyxl

---

📊 Dataset Usage

• Official datasets provided by NeuroLogic '26 organizers
• Training: with_label files used for model training
• Evaluation: no_label files used for predictions
• Evaluation method: 90/10 train-validation split, seed=42

---

✅ Pre-trained Models

Model	Challenge	License
vinai/bertweet-base	Challenge 1	MIT
roberta-base	Challenge 2	MIT
xlm-roberta-base	Challenge 3	MIT

All models publicly available on HuggingFace Hub. Use of pre-trained models explicitly permitted per NeuroLogic '26 rules.