ETSP — Multilingual Counterspeech Generation

This repository contains the code for a semester project on Multilingual Counterspeech Generation, conducted as part of a university course and inspired by the Shared Task on Multilingual Counterspeech Generation (MCG-COLING 2025).

The project investigates how large language models fine-tuned with parameter-efficient methods perform across high- and low-resource languages, with a focus on multilingual generalization and training dynamics.

---

Project Overview

• Task: Generate respectful and factual counter-narratives in response to hate speech
• Languages: English, Italian, Spanish, Basque
• Model: LLaMA-2-7B
• Fine-tuning: LoRA + 4-bit quantization (NF4)
• Evaluation: BLEU, ROUGE-L, BERTScore

---

Repository Structure

• Main training and evaluation notebook: Counterspeech.py
• Knowledge selection: kwnoledge_selection_gpt4o.py
• Dataset used for knowledge selection part: train_selected.csv, validation_selected.csv
• Project documentation: README.md

Dataset

This project uses the dataset released for the MCG-COLING Shared Task, consisting of:

• 596 Hate Speech – Counter Narrative pairs per language
• 5 background knowledge sentences per instance
• Splits:
  • Train: 396
  • Development: 100
  • Test: 100

Dataset source:

• Hugging Face: LanD-FBK/ML_MTCONAN_KN

---

Methodology

Model

• Base model: meta-llama/Llama-2-7b-hf
• Objective: Causal Language Modeling

Fine-Tuning

• LoRA rank: 8
• LoRA alpha: 64
• Dropout: 0.05
• Quantization: 4-bit NF4 using bitsandbytes
• Frameworks: Hugging Face Transformers, PEFT

Each language is fine-tuned independently using identical hyperparameters to enable fair cross-lingual comparison.

---

Evaluation

Model performance is evaluated using the following automatic metrics:

• BLEU
• ROUGE-L
• BERTScore

Training dynamics are monitored using TensorBoard to analyze convergence behavior across languages.

---

How to Run

1. Install dependencies

pip install torch transformers datasets peft bitsandbytes accelerate evaluate bert_score rouge_score

2. Run training and evaluation

Open and run the Python file:

Counterspeech.py

A GPU with sufficient VRAM is strongly recommended.