Qwen3-8B fine-tune lost its thinking mode because the template taught empty `<think>` blocks

MeldhLLC · reddit · 2026-07-26

A Qwen3-8B fine-tune silently lost its thinking mode because the training template taught the model to emit empty <think> blocks.

The author says two fine-tunes on different versions of the same corpus produced different outcomes: one still responded to the thinking toggle, while the other always returned an empty thinking field. Standard loss metrics did not catch the regression.

Key findings:

Qwen3’s chat template renders non-thinking turns with an empty <think> block.
Reusing that template for answer-only fine-tuning can train the model to prefer empty reasoning blocks.
A cited paper saw the same issue: valid-looking reasoning blocks can pass checks even when the reasoning is wrong or empty.
On GSM8K, the model produced valid reasoning in 96.5% of outputs when targets omitted think tags, 58.2% when empty think blocks were included, and 100% when those empty blocks were excluded from loss.
In their own false-premise benchmark, the build that lost thinking fabricated more often than the base or thinking-capable versions.

They also found an Ollama template pitfall: the imported GGUF’s embedded template overrode the Modelfile template, so the prompt never reached the model as expected. Once they rendered prompts themselves, the same weights deliberated on 20/20 test prompts.

Original post →

Qwen3-8B fine-tune lost its thinking mode because the template taught empty `<think>` blocks

More from Models