Browse Teach Curate

Fine-Tuning Compact Language Models for Multiple-Choice Question Answering

How a small language model such as Phi-3 can be adapted to answer multiple-choice questions reliably by combining supervised fine-tuning on a standardized MCQ dataset with careful prompt design. Students learn why naive completion prompts induce positional bias (the model always picking the last option), how structured Alpaca-style prompts and option standardization fix it, and how perplexity, accuracy, and F1 quantify the improvement in resource-constrained settings.

Terminator Text

Fine-Tuning PHI-3 for Multiple-Choice Question Answering: Methodology, Results, and Challenges

This paper fine-tunes Microsoft's Phi-3 (1.3B-parameter) compact language model to answer multiple-choice questions, using the TruthfulQA dataset preprocessed to a consistent option format. It combin…