xwal3/nllb_lora_en_am_finetuned

This project demonstrates a LoRA-based fine-tuning workflow on the facebook/nllb-200-distilled-600M model using "habtew/english-amharic-translation" dataset for English → Amharic translation. The goal was to experiment with parameter-efficient fine-tuning and build a small, reproducible pipeline for evaluation and sample translations. # LoRA finetuning NLLB en to am Model. Model training and fine tuning using LoRA on facebook/nllb-200-distilled-600M model for English to Amharic translation. ## 📂 Models and Datasets. - NLLB-600M-disttled-600M - Model used to train. - Habtew-english-to-amharic-translation - Dataset for training. ## 🛠 Workflow Overview + Dataset Preparation + Sampled subsets from 173k train / 21.6k validation / 43.1k test to 10k train / 2.5k validation & test . + Tokenized datasets for Seq2Seq model input. + Sanity Check + Ran a 100-row RAG pipeline sanity check. + Ensured model outputs are consistent before full training. + LoRA Fine-Tuning + Reduced trainable parameters drastically while retaining most of the base model knowledge. + Trained on 10k samples with 3 epochs. + Batch Evaluation + Generated predictions on test dataset. + Calculated CHRF++ scores for quality evaluation. + Saving Results + Saved trained LoRA adapters and tokenizer. + Saved evaluation outputs and sampled datasets as JSON. # 📊 Results The fine-tuned model demonstrated the ability to learn from the dataset and produced reasonable translations in many cases. Below are examples: ```json { "source": "who gives jesus the authority to conquer?", "reference": "ኢየሱስ ድል እንዲያደርግ ስልጣን የሰጠው ማን ነው?", "prediction": "ኢየሱስ ድል እንዲያደርግ ሥልጣን የሰጠው ማን ነው?", "chrf++": 83.23655149198628 } ``` Here, the model produced an accurate translation that closely matches the reference. ```json { "source": "\"If gun control worked, Washington, D.C., would be the beacon.", "reference": "ይሁን እንጂ የዩናይትድ እስቴትስ ግድያ የበዛበት ዋና ከተማ ናት” ብለዋል።", "prediction": "\"የጦር መሣሪያ ቁጥጥር ቢሰራ ዋሽንግተን ዲሲ የጦር መሣሪያ መቆጣጠሪያ ማዕከል ትሆን ነበር።", "chrf++": 4.396984924623116 } ``` In this case, the prediction diverged significantly from the reference, resulting in a very low CHRF++ score. Overall the model has learned with sample dataset and few epocsh and the overall valuation is: | Metric | Score | | ------ | ------ | | Corpus CHRF++ (LoRA fine-tuned) | 33.03 | …