github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-06 16:11:34 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #15798] gemma4-64k chat-template special tokens (<|tool_call|>, <|"|>, <|channel|>) leak into OpenAI-compat tool-calling output #72125

New Issue
Closed
opened 2026-05-05 03:30:55 -05:00 by GiteaMirror · 1 comment
GiteaMirror commented 2026-05-05 03:30:55 -05:00
Owner
Copy Link

Originally created by @mons-bot on GitHub (Apr 24, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15798

Observed

Running gemma4-64k:latest through Ollama's OpenAI-compat endpoint (/v1/chat/completions with tool schemas), the model emits tool calls as plain text inside the assistant message — often inside a thinking block — with Gemma chat-template special tokens leaked verbatim (<|tool_call|>, <|"|>, <|channel|>, <|tool_response|>). finish_reason returns stop, so the harness ends the turn as if the model had spoken a normal reply, and no tool call is dispatched.

Ollama version

0.21.1. Prior related issues (#15241, #15315) claim a fix in 0.20.6 but the behavior is still reproducible on 0.21.1.

Pattern

  • String arguments wrapped in <|"|>...<|"|> instead of real quotes.
  • call: prefix where a structured tool-call should be.
  • Surrounding <|tool_call|> / <|channel|> template tokens emitted as plain text inside content.

Workaround

None. Swapping model family is the only reliable option.

Closing: will redo with a generic minimal repro if needed.

Originally created by @mons-bot on GitHub (Apr 24, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15798 ### Observed Running `gemma4-64k:latest` through Ollama's OpenAI-compat endpoint (`/v1/chat/completions` with tool schemas), the model emits tool calls as plain text inside the assistant message — often inside a `thinking` block — with Gemma chat-template special tokens leaked verbatim (`<|tool_call|>`, `<|"|>`, `<|channel|>`, `<|tool_response|>`). `finish_reason` returns `stop`, so the harness ends the turn as if the model had spoken a normal reply, and no tool call is dispatched. ### Ollama version 0.21.1. Prior related issues (#15241, #15315) claim a fix in 0.20.6 but the behavior is still reproducible on 0.21.1. ### Pattern - String arguments wrapped in `<|"|>...<|"|>` instead of real quotes. - `call:` prefix where a structured tool-call should be. - Surrounding `<|tool_call|>` / `<|channel|>` template tokens emitted as plain text inside `content`. ### Workaround None. Swapping model family is the only reliable option. Closing: will redo with a generic minimal repro if needed.
GiteaMirror commented 2026-05-05 03:30:57 -05:00
Author
Owner
Copy Link

@fabiopolimeni commented on GitHub (Apr 30, 2026):

This happens and makes the gemma4 family unusable for agentic workflows that hit OpenAI compatible API (almost all by default use it).

<!-- gh-comment-id:4350863772 --> @fabiopolimeni commented on GitHub (Apr 30, 2026): This happens and makes the gemma4 family unusable for agentic workflows that hit OpenAI compatible API (almost all by default use it).
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
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#72125
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?