Skip to main content

Chat-Template Debugger

If your training tokens and inference tokens differ by one space, the model quietly gets worse.

Key Insight​

A chat template wraps each turn in special tokens (system, user, assistant). The model was fine-tuned on one exact byte pattern; rendering it by hand and byte-comparing surfaces any mismatch.

Why This Matters​

A single off-by-one space between training and inference is one of the most common — and hardest to spot — causes of a fine-tuned model silently underperforming.