fix(chat_templates): extract reasoning_content before reassigning content in qwen3_5 template

[Anai-Guo]

Summary

Fixes #3724.

The bundled qwen3_5.jinja template parses inline <think>...</think> reasoning incorrectly for assistant messages. In the else branch that handles inline reasoning, the two {%- set %} statements were in the wrong order relative to the official Qwen3.5 template: content was mutated before reasoning_content was extracted from it.

Both assignments read content, so they are order-dependent:

{%- set content = content.split('</think>')[-1].lstrip('\n') %}
{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}

After the first line, content no longer contains </think> (or <think>), so the second line's content.split('</think>')[0].split('<think>')[-1] just returns the already-truncated answer. The result: the reasoning trace is silently dropped and the answer is duplicated into the <think> block of the rendered prompt.

Fix

Swap the two statements so reasoning_content is extracted from the original content before content is reassigned, matching the official Qwen3.5 template:

{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
{%- set content = content.split('</think>')[-1].lstrip('\n') %}

Test plan

• Render an assistant turn whose content contains an inline <think>...</think> block (and no separate message.reasoning_content field) and confirm the reasoning text lands in the <think> block and the answer follows after </think>.

🤖 Generated with Claude Code

Summary by CodeRabbit

• Bug Fixes
  • Fixed parsing and display of reasoning content in model responses to ensure accurate extraction.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 6b8a1088-7722-45e9-a615-a80b37adc4b9

📥 Commits

Reviewing files that changed from the base of the PR and between 22bcb9a and cee969b.

📒 Files selected for processing (1)

• src/axolotl/utils/chat_templates/templates/qwen3_5.jinja

---

📝 Walkthrough

Walkthrough

The PR fixes a bug in the Qwen3.5 chat template where inline <think>...</think> reasoning blocks were being silently dropped. Two variable assignments in the assistant message rendering logic are reordered so reasoning_content is extracted before content is mutated.

Changes

Qwen3.5 Template Reasoning Parsing Fix

Layer / File(s)	Summary
Reorder reasoning extraction before content mutation src/axolotl/utils/chat_templates/templates/qwen3_5.jinja	Lines 96–97 are reordered: reasoning_content is now computed from the original content before content is reassigned to the portion after </think>, fixing the inline reasoning block extraction logic that was previously dropping reasoning when message.reasoning_content is not provided as a separate field.

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

Suggested reviewers

• winglian

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title accurately describes the main fix: reordering two Jinja set statements to extract reasoning_content before reassigning content in the qwen3_5 template.
Linked Issues check	✅ Passed	The PR correctly implements the primary fix from issue #3724: swapping the two set statements so reasoning_content is extracted before content is reassigned, resolving the bug where inline reasoning was lost.
Out of Scope Changes check	✅ Passed	The PR makes only the minimal targeted fix to qwen3_5.jinja (1 line added, 1 line removed) that directly addresses the statement reordering issue in issue #3724, with no extraneous changes.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[ved1beta]

good catch

[NanoCode012]

Ooh, this was the same comment I had on your PR @ved1beta . I think you can choose whichever to merge first