Writing an LLM from scratch -- Updated instruction fine-tuning results

A developer tested instruction fine-tuning on multiple GPT-2-style language models to evaluate real-world usefulness beyond technical loss metrics. Despite expectations that lower loss would correlate with better instruction-following, results showed surprising inconsistencies, with some high-performing models scoring poorly and models trained on educational data outperforming technically superior ones. The findings suggest that a model's position in the loss landscape doesn't guarantee good performance after instruction fine-tuning, indicating that chasing lower loss alone may not produce the most useful models.