github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 00:22:43 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #6870] Nemotron-mini Tool Calling not using the tools API 50% of the time #30100

New Issue
Closed
opened 2026-04-22 09:34:02 -05:00 by GiteaMirror · 7 comments
GiteaMirror commented 2026-04-22 09:34:02 -05:00
Owner
Copy Link

Originally created by @whogben on GitHub (Sep 19, 2024).
Original GitHub issue: https://github.com/ollama/ollama/issues/6870

What is the issue?

When Nemotron-mini tries to call tools, it often returns this XML wrapping a JSON tool call string, rather than triggering the Ollama tools API - see this message payload:

"message": {
    "content": " <toolcall> {\"name\": \"weather_lookup\", \"arguments\": {\n  \"temp_units\": \"celsius\",\n  \"zip_codes\": [\n    12345\n  ]\n}} </toolcall>",
    "role": "assistant"
},

However, sometimes it does use the tool calling API, (example correct payload):

"message": {
		"content": "",
		"role": "assistant",
		"tool_calls": [
			{
				"function": {
					"arguments": {
						"temp_units": "celsius",
						"zip_codes": [
							"12345"
						]
					},
					"name": "weather_lookup"
				}
			}
		]
	}

What I have noticed is that it either succeeds right away, or it gets locked into repeatedly making the unrecognized tool call indefinitely - eg, once it has it's own wrong-example in the chat history, it persists in using that format indefinitely.
I'm not sure what the recommendation would be here - normally I'd expect one should edit the template, but since it works about 50% of the time and fails about 50% of the time, I'm not sure which style to represent in the template.

OS

Windows

GPU

Nvidia

CPU

Intel

Ollama version

0.3.11

Originally created by @whogben on GitHub (Sep 19, 2024). Original GitHub issue: https://github.com/ollama/ollama/issues/6870 ### What is the issue? When [Nemotron-mini](https://ollama.com/library/nemotron-mini) tries to call tools, it often returns this XML wrapping a JSON tool call string, rather than triggering the Ollama tools API - see this message payload: ``` "message": { "content": " <toolcall> {\"name\": \"weather_lookup\", \"arguments\": {\n \"temp_units\": \"celsius\",\n \"zip_codes\": [\n 12345\n ]\n}} </toolcall>", "role": "assistant" }, ``` However, sometimes it does use the tool calling API, (example correct payload): ``` "message": { "content": "", "role": "assistant", "tool_calls": [ { "function": { "arguments": { "temp_units": "celsius", "zip_codes": [ "12345" ] }, "name": "weather_lookup" } } ] } ``` What I have noticed is that it either succeeds right away, or it gets locked into repeatedly making the unrecognized tool call indefinitely - eg, once it has it's own wrong-example in the chat history, it persists in using that format indefinitely. I'm not sure what the recommendation would be here - normally I'd expect one should edit the template, but since it works about 50% of the time and fails about 50% of the time, I'm not sure which style to represent in the template. ### OS Windows ### GPU Nvidia ### CPU Intel ### Ollama version 0.3.11
GiteaMirror added the bug label 2026-04-22 09:34:02 -05:00
GiteaMirror commented 2026-04-22 09:34:04 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 19, 2024):

What's the code you are using to make the request that responds with <toolcall>?

<!-- gh-comment-id:2359956594 --> @rick-github commented on GitHub (Sep 19, 2024): What's the code you are using to make the request that responds with `<toolcall>`?
GiteaMirror commented 2026-04-22 09:34:04 -05:00
Author
Owner
Copy Link

@whogben commented on GitHub (Sep 19, 2024):

Identical payload each time, I'm using the REST API and the results vary between those two examples above with no code changes. Here's the payload:

{
	"messages": [
		{
			"content": "What is the weather in Example State, zip 12345, in degrees Celsius?",
			"role": "user"
		}
	],
	"model": "nemotron-mini:latest",
	"stream": false,
	"tools": [
		{
			"function": {
				"description": "Use this tool to get the weather.",
				"name": "weather_lookup",
				"parameters": {
					"properties": {
						"temp_units": {
							"description": "What units to use for the temperature.",
							"enum": [
								"fahrenheit",
								"celsius",
								"kelvin"
							],
							"type": "string"
						},
						"zip_codes": {
							"description": "An array of zip codes to get the weather for.",
							"items": {
								"description": " A 5 digit USA zipcode string",
								"type": "string"
							},
							"type": "array"
						}
					},
					"required": [
						"zip_codes"
					],
					"type": "object"
				}
			},
			"type": "function"
		}
	]
}

(I've redacted my actual state and zip code)
I've also tested with Hermes 3 and Llama-3.1-groq-tool-calling-preview (maybe 5 tries each, both worked as usual) - so no code changes, but swapping the model does change the result.

<!-- gh-comment-id:2361370721 --> @whogben commented on GitHub (Sep 19, 2024): Identical payload each time, I'm using the REST API and the results vary between those two examples above with no code changes. Here's the payload: ``` { "messages": [ { "content": "What is the weather in Example State, zip 12345, in degrees Celsius?", "role": "user" } ], "model": "nemotron-mini:latest", "stream": false, "tools": [ { "function": { "description": "Use this tool to get the weather.", "name": "weather_lookup", "parameters": { "properties": { "temp_units": { "description": "What units to use for the temperature.", "enum": [ "fahrenheit", "celsius", "kelvin" ], "type": "string" }, "zip_codes": { "description": "An array of zip codes to get the weather for.", "items": { "description": " A 5 digit USA zipcode string", "type": "string" }, "type": "array" } }, "required": [ "zip_codes" ], "type": "object" } }, "type": "function" } ] } ``` (I've redacted my actual state and zip code) I've also tested with Hermes 3 and Llama-3.1-groq-tool-calling-preview (maybe 5 tries each, both worked as usual) - so no code changes, but swapping the model does change the result.
GiteaMirror commented 2026-04-22 09:34:05 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Sep 21, 2024):

I ran the query 200 times and apart from differences in argument quoting, it worked as expected:

$ for i in {1..200} ; do curl -s localhost:11434/api/chat -d "$(cat 6870.prompt)" | jq -c .message.tool_calls ; done | sort | uniq -c
     51 [{"function":{"name":"weather_lookup","arguments":{"temp_units":"celsius","zip_codes":["12345"]}}}]
    149 [{"function":{"name":"weather_lookup","arguments":{"temp_units":"celsius","zip_codes":[12345]}}}]

I pulled the model this morning specifically for testing, I see that the model is marked as updated yesterday, and you filed the ticket two days ago, so perhaps a model update has taken care of the issue.

<!-- gh-comment-id:2365046886 --> @rick-github commented on GitHub (Sep 21, 2024): I ran the query 200 times and apart from differences in argument quoting, it worked as expected: ```console $ for i in {1..200} ; do curl -s localhost:11434/api/chat -d "$(cat 6870.prompt)" | jq -c .message.tool_calls ; done | sort | uniq -c 51 [{"function":{"name":"weather_lookup","arguments":{"temp_units":"celsius","zip_codes":["12345"]}}}] 149 [{"function":{"name":"weather_lookup","arguments":{"temp_units":"celsius","zip_codes":[12345]}}}] ``` I pulled the model this morning specifically for testing, I see that the model is marked as updated yesterday, and you filed the ticket two days ago, so perhaps a model update has taken care of the issue.
GiteaMirror commented 2026-04-22 09:34:06 -05:00
Author
Owner
Copy Link

@whogben commented on GitHub (Sep 23, 2024):

Yup working for me too after I pulled the latest, thanks for testing!

<!-- gh-comment-id:2369140528 --> @whogben commented on GitHub (Sep 23, 2024): Yup working for me too after I pulled the latest, thanks for testing!
GiteaMirror commented 2026-04-22 09:34:06 -05:00
Author
Owner
Copy Link

@tripolskypetr commented on GitHub (Jan 2, 2025):

After the update it still spamming <toolcall> {\"type\": \"function\", \"arguments\": {\n \"name\": \"navigate_to_sales_agent_tool\"\n}} </toolcall> to the text chat instead of function execution. Please downgrade the model

<!-- gh-comment-id:2567601639 --> @tripolskypetr commented on GitHub (Jan 2, 2025): After the update it still spamming `<toolcall> {\"type\": \"function\", \"arguments\": {\n \"name\": \"navigate_to_sales_agent_tool\"\n}} </toolcall>` to the text chat instead of function execution. Please downgrade the model
GiteaMirror commented 2026-04-22 09:34:07 -05:00
Author
Owner
Copy Link

@rick-github commented on GitHub (Jan 2, 2025):

Example code?

<!-- gh-comment-id:2567648083 --> @rick-github commented on GitHub (Jan 2, 2025): Example code?
GiteaMirror commented 2026-04-22 09:34:08 -05:00
Author
Owner
Copy Link

@tripolskypetr commented on GitHub (Jan 5, 2025):

Got the steps to reproduce in https://github.com/ollama/ollama/issues/8287#issuecomment-2571758553

<!-- gh-comment-id:2571758993 --> @tripolskypetr commented on GitHub (Jan 5, 2025): Got the steps to reproduce in https://github.com/ollama/ollama/issues/8287#issuecomment-2571758553
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 bug
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#30100
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?