github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 08:02:14 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #9627] removing <think> tag as an option #68338

New Issue
Closed
opened 2026-05-04 13:15:41 -05:00 by GiteaMirror · 5 comments
GiteaMirror commented 2026-05-04 13:15:41 -05:00
Owner
Copy Link

Originally created by @Shahin-rmz on GitHub (Mar 10, 2025).
Original GitHub issue: https://github.com/ollama/ollama/issues/9627

Hello, Ollama made my life much easier. Thanks a lot.
For the new generation of reasoning models that they have thing tag, is there a way to delete/remove the
</think> tag if we want? if not it will be cool idea to have that option both in API and direct usage one. thanks a lot.

Originally created by @Shahin-rmz on GitHub (Mar 10, 2025). Original GitHub issue: https://github.com/ollama/ollama/issues/9627 Hello, Ollama made my life much easier. Thanks a lot. For the new generation of reasoning models that they have thing tag, is there a way to delete/remove the ```</think>``` tag if we want? if not it will be cool idea to have that option both in API and direct usage one. thanks a lot.
GiteaMirror added the feature request label 2026-05-04 13:15:41 -05:00
GiteaMirror commented 2026-05-04 13:15:43 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 10, 2025):

The <think> tags are a function of the model and there's currently no mechanism for removing them from the token stream generated by the model. The client usually does this, for example OpenWebUI detects them and hides them behind a spinner. API clients can detect the opening tag and drop output until the closing tag is detected. If you are using the deepseek distillate models, then you can use the base model (qwen2.5 or llama3.1) and get output without the reasoning steps.

<!-- gh-comment-id:2711527607 --> @rick-github commented on GitHub (Mar 10, 2025): The `<think>` tags are a function of the model and there's currently no mechanism for removing them from the token stream generated by the model. The client usually does this, for example OpenWebUI detects them and hides them behind a spinner. API clients can detect the opening tag and drop output until the closing tag is detected. If you are using the deepseek distillate models, then you can use the base model (qwen2.5 or llama3.1) and get output without the reasoning steps.
GiteaMirror commented 2026-05-04 13:15:43 -05:00
Author
Owner
Copy Link

@pdevine commented on GitHub (Mar 12, 2025):

@Shahin-rmz I could see us potentially putting something into the CLI to hide it, but I'm not sure it's a great idea for the API. As @rick-github mentioned, usually the client would filter these out.

<!-- gh-comment-id:2716129534 --> @pdevine commented on GitHub (Mar 12, 2025): @Shahin-rmz I could see us potentially putting something into the CLI to hide it, but I'm not sure it's a great idea for the API. As @rick-github mentioned, usually the client would filter these out.
GiteaMirror commented 2026-05-04 13:15:44 -05:00
Author
Owner
Copy Link

@allenporter commented on GitHub (Mar 15, 2025):

Usually thinking steps are expose separate in the API. The client shouldn't need to implement its own way to strip internal model tokens, just like any other internal model token. It can show or filter by processing the thinking steps in the API. This is how it works in other cloud APIs (Anthropic, google, etc) as well

<!-- gh-comment-id:2726748005 --> @allenporter commented on GitHub (Mar 15, 2025): Usually thinking steps are expose separate in the API. The client shouldn't need to implement its own way to strip internal model tokens, just like any other internal model token. It can show or filter by processing the thinking steps in the API. This is how it works in other cloud APIs (Anthropic, google, etc) as well
GiteaMirror commented 2026-05-04 13:15:44 -05:00
Author
Owner
Copy Link

@Kwisss commented on GitHub (Mar 19, 2025):

So if I understand correctly, there is no way for Ollama to distinct the thinking from the actual response in a model like qwq?
like returning it as a different type or separated from content?

I did this to separate the response but it could probably be done a lot better:

def generate():
	nonlocal assistant_response
	for line in response.iter_lines():
		if line:
			chunk = json.loads(line.decode('utf-8'))
			if 'message' in chunk and 'content' in chunk['message']:
				content = chunk['message']['content']
				assistant_response += content
				yield f"data: {json.dumps({'content': content})}\n\n"
			if chunk.get('done', False):
				# Split into thinking and content
				if '</think>' in assistant_response:
					thinking_response, content_response = assistant_response.split('</think>', 1)
				else:
					thinking_response = ''
					content_response = assistant_response
				# Store in conversation history
				conversations[session_id].append({
					"role": "assistant",
					"content": content_response.strip(),
					"thinking": thinking_response.strip()
				})
				break
	yield f"data: {json.dumps({'done': True})}\n\n"

return app.response_class(generate(), mimetype='text/event-stream')
<!-- gh-comment-id:2737072079 --> @Kwisss commented on GitHub (Mar 19, 2025): So if I understand correctly, there is no way for Ollama to distinct the thinking from the actual response in a model like qwq? like returning it as a different type or separated from content? I did this to separate the response but it could probably be done a lot better: ``` def generate(): nonlocal assistant_response for line in response.iter_lines(): if line: chunk = json.loads(line.decode('utf-8')) if 'message' in chunk and 'content' in chunk['message']: content = chunk['message']['content'] assistant_response += content yield f"data: {json.dumps({'content': content})}\n\n" if chunk.get('done', False): # Split into thinking and content if '</think>' in assistant_response: thinking_response, content_response = assistant_response.split('</think>', 1) else: thinking_response = '' content_response = assistant_response # Store in conversation history conversations[session_id].append({ "role": "assistant", "content": content_response.strip(), "thinking": thinking_response.strip() }) break yield f"data: {json.dumps({'done': True})}\n\n" return app.response_class(generate(), mimetype='text/event-stream') ```
GiteaMirror commented 2026-05-04 13:15:45 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Mar 19, 2025):

https://github.com/ollama/ollama/issues/8528#issuecomment-2736910261

<!-- gh-comment-id:2737081427 --> @rick-github commented on GitHub (Mar 19, 2025): https://github.com/ollama/ollama/issues/8528#issuecomment-2736910261
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
amd
api
app
bug
build
cli
cloud
compatibility
context-length
create
docker
documentation
embeddings
feature request
feedback wanted
good first issue
gpt-oss
gpu
harmony
help wanted
image
install
intel
js
launch
linux
macos
memory
mlx
model
needs more info
networking
nvidia
ollama.com
performance
pull-request
Mirrored from GitHub Pull Request
 python
question
registry
rendering
thinking
tools
top
vulkan
windows
wsl
No Label feature request
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
GiteaMirror ninjasurge
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: github-starred/ollama#68338
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?