Bekyy/Amharic-word-generation-GenAI-

# Amharic-word-generation-GenAI- ## 📌 Project Overview This project focuses on generating antonyms for Amharic words using a pretrained transformer model. The model is fine-tuned to learn word relationships, specifically antonyms, leveraging T5 (Text-to-Text Transfer Transformer) for sequence generation. ## 🚀 Features * Fine-tuned facebook/m2m100_418M for Amharic antonym generation. * Training on a structured dataset containing Amharic word pairs labeled as antonyms. * Evaluation using BLEU score and accuracy metrics. * Deployment-ready implementation using Hugging Face Transformers. ## 📂 Dataset * The dataset consists of three columns: - word1 - The target Amharic word. - word2 - The corresponding antonym. * The data is preprocessed and tokenized using Hugging Face's Tokenizer. ## 🏗 Model Training * The model is trained using T5ForConditionalGeneration with the following setup: ```python from transformers import AutoModelForSeq2SeqLM, Seq2SeqTrainer, Seq2SeqTrainingArguments, EarlyStoppingCallback model = AutoModelForSeq2SeqLM.from_pretrained("facebook/m2m100_418M") training_args = Seq2SeqTrainingArguments( output_dir="./m2m-amharic-antonym-augmented", # Directory to save the model evaluation_strategy="epoch", # Evaluate at the end of each epoch learning_rate=2e-5, # Learning rate per_device_train_batch_size=16, # Training batch size per_device_eval_batch_size=16, # Evaluation batch size num_train_epochs=100, # Number of epochs weight_decay=0.01, # Weight decay for regularization predict_with_generate=True, # Allow prediction generation # logging_dir="./logs", # Directory for logs logging_steps=10, # Log every 10 steps push_to_hub=True, # Enable pushing to Hugging Face hub report_to="tensorboard", save_strategy="epoch", save_total_limit=1, # Limit on the number of saved checkpoints load_best_model_at_end=True, ) ear …