How Chat Templates Actually Work in Supervised Fine-Tuning

When fine-tuning language models, the structured 'messages' list used in Python never reaches the model — only the rendered text string does. Chat templates are Jinja2 programs stored in tokenizer configuration files that flatten conversation structure into a single token sequence. Roles like 'user' and 'assistant' are represented as actual vocabulary tokens, not metadata fields, meaning the model learns turn-taking purely from token patterns. A key distinction exists between training, which renders the full conversation, and inference, which appends a generation prompt — and both must produce identical prefixes to avoid mismatches. Common pitfalls include accidentally adding duplicate special tokens and overlooking how whitespace stripping in Jinja2 tags directly affects the final token sequence.