angelin-mary/Off-Target-Mitigation-in-NLLB-

A implementation of Meta’s “No Language Left Behind” (NLLB) machine translation model enhanced with Target Language Prediction (TLP) for improved performance on low-resource language pairs. Includes training scripts and evaluation pipelines. # Off-Target Mitigation in NLLB This repository provides scripts and models to evaluate and improve the performance of the NLLB (No Language Left Behind) multilingual translation system using **Target Language Prediction (TLP)** to mitigate off-target translations. It includes training and evaluation pipelines for both baseline and fine-tuned models. --- ## Repository Structure ### 🔹 Translation Evaluation | Script | Description | |--------|-------------| | `NLLB_Baseline_test_eng.py` | Evaluate baseline NLLB model on English → many translation. | | `NLLB_Baseline_test_mal.py` | Evaluate baseline NLLB model on Malayalam → many translation. | | `NLLB_Baseline_test_spa.py` | Evaluate baseline NLLB model on Spanish → many translation. | | `NLLB_TLP_test_eng.py` | Evaluate fine-tuned TLP model on English → many. | | `NLLB_TLP_test_mal.py` | Evaluate fine-tuned TLP model on Malayalam → many. | | `NLLB_TLP_test_spa.py` | Evaluate fine-tuned TLP model on Spanish → many. | ### 🔹 Language Identification | Script | Description | |--------|-------------| | `languageIdentifier_Train.py` | Trains a language identifier using FLORES-200 and XLM-RoBERTa. | | `languageIdentifierEvaluation.py` | Evaluates the trained language identification model. | ### 🔹 Correlation Analysis | Script | Description | |--------|-------------| | `correlationAnalysis.py` | Analyzes correlation between off-target rates and spBLEU. | | `correlation_visualisation.py` | Visualizes off-target vs BLEU score trends. | ### 🔹 TLP Training | Script | Description | |--------|-------------| | `NLLB_TLP_training.py` | Fine-tunes the NLLB model with Target Language Prediction (TLP). | --- ## Usage Run the provided `run.sh` script to install all required dependencies using Python 3.8.