Why I Pivoted from a Base Model to a Chat Model
The first version of my Llama2 code-generation setup used a base model strategy because it looked simple and lightweight. In practice, the generated code had inconsistent structure and weaker task alignment, which made Pass@k evaluation noisy.
The pivot to a chat-model format forced me to clarify prompt templates and training instruction pairs. That produced outputs that were easier to execute, easier to compare, and easier to score. It was not just a model swap; it was a pipeline-quality decision.
The useful lesson was that architecture choices are less about trend-following and more about reducing uncertainty inside evaluation. If you cannot trust your evaluation signal, you cannot trust your next training decision.
Email me