--- license: apache-2.0 datasets: - rajpurkar/squad_v2 metrics: - precision - f1 - recall - squad_v2 - meteor - bleu - rouge - exact_match base_model: - meta-llama/Llama-3.2-1B - google/gemma-2-2b-it library_name: transformers tags: - llama - sqaud - fine - tuned --- 1. Overview This repository highlights the fine-tuning of the Llama-3.2-1B model on the SQuAD (Stanford Question Answering Dataset) dataset. The task involves training the model to accurately answer questions based on a given context passage. Fine-tuning the pre-trained Llama model aligns it with the objectives of extractive question-answering. 2. Model Information Model Used: meta-llama/Llama-3.2-1B Pre-trained Parameters: The model contains approximately 1.03 billion parameters, verified during setup and matching official documentation. Fine-tuned Parameters: The parameter count remains consistent with the pre-trained model, as fine-tuning only updates task-specific weights. 3. Dataset and Task Details Dataset: SQuAD The Stanford Question Answering Dataset (SQuAD) is a benchmark dataset designed for extractive question-answering tasks. It contains passages with corresponding questions and answer spans extracted directly from the text. Task Objective Given a passage and a question, the model is trained to identify the correct span of text in the passage that answers the question. 4. Fine-Tuning Approach Train-Test Split: An 80:20 split was applied to the dataset, ensuring a balanced distribution of passages and questions in the train and test subsets. Stratified sampling was used, with a seed value of 1 for reproducibility. Tokenization: Context and question pairs were tokenized with padding and truncation to ensure uniform input lengths (maximum 512 tokens). Model Training: Fine-tuning was conducted over three epochs with a learning rate of 3e-5. Gradient accumulation and early stopping were used to enhance training efficiency and prevent overfitting. Hardware: Training utilized GPU acceleration to handle the large model size and complex token sequences efficiently. 5. Results and Observations Zero-shot vs. Fine-tuned Performance: Without fine-tuning, the pre-trained Llama model demonstrated limited ability to answer questions accurately. Fine-tuning significantly improved the model’s performance on metrics such as F1 score, exact match, and ROUGE. Fine-tuning Benefits: Training on the SQuAD dataset equipped the model with a deeper understanding of context and its relationship to specific queries, enhancing its ability to extract precise answer spans. Model Parameters: The parameter count remained unchanged during fine-tuning, underscoring that performance improvements stemmed from the optimization of existing weights rather than structural changes. 6. How to Use the Fine-Tuned Model Install Necessary Libraries: pip install transformers datasets Load the Fine-Tuned Model: from transformers import AutoTokenizer, AutoModelForQuestionAnswering model_name = "/squad-llama-finetuned" tokenizer = AutoTokenizer.from_pretrained(model_name) model = AutoModelForQuestionAnswering.from_pretrained(model_name) Make Predictions: context = "Llama is a model developed by Meta AI designed for natural language understanding tasks." question = "Who developed Llama?" inputs = tokenizer(question, context, return_tensors="pt", truncation=True, padding=True) outputs = model(**inputs) start_idx = outputs.start_logits.argmax() end_idx = outputs.end_logits.argmax() answer = tokenizer.decode(inputs["input_ids"][0][start_idx:end_idx + 1]) print(f"Predicted Answer: {answer}") 7. Key Takeaways Fine-tuning Llama on SQuAD equips it with the ability to handle extractive question-answering tasks with high accuracy and precision. The parameter count of the model does not change during fine-tuning, highlighting that performance enhancements are derived from weight updates rather than architectural modifications. The comparison between zero-shot and fine-tuned performance demonstrates the necessity of task-specific training to achieve state-of-the-art results. 8. Acknowledgments Hugging Face for providing seamless tools for model fine-tuning and evaluation. Stanford Question Answering Dataset for serving as a robust benchmark for extractive QA tasks.