Primitive of the day · 2026-09-28

finetune-chat-template-exact-match

A fine-tuned GGUF that loads fine but generates garbage/repetition almost always has a chat-template mismatch — the Ollama Modelfile TEMPLATE must be character-for-character the template the training data used; prevent the whole class by using tokenizer.apply_chat_template() at data-prep time.

When this applies

A fine-tuned model imported into Ollama produces repetition, wrong format, or refusals even though the GGUF loads without error.

Preconditions

• The chat template used to format the training data is known (which token pattern family)

Do this

Set the Modelfile TEMPLATE to exactly the training-time template — per family: Llama-3 <|start_header_id|>/<|eot_id|>, Mistral/Mixtral [INST], Qwen ChatML <|im_start|>/<|im_end|>, Phi-3 <|user|>/<|end|>, Gemma <start_of_turn>/<end_of_turn>; going forward, format training data with apply_chat_template so data, tokenizer, and Modelfile agree.

What you should see

Coherent, correctly formatted generations after ollama create.

How it fails if ignored

The doc's 'MOST COMMON failure mode': garbage output with a perfectly valid GGUF, typically misdiagnosed as a bad fine-tune.

Do not use when

Base models pulled from the Ollama registry (templates ship correct); actual GGUF corruption (re-fuse instead — 'invalid model' at create time).

Kind: gotcha-fix. Part of the skill MLX lm fuse and GGUF export. Free to reuse in your own agent skills.