github-starred/ollama
2
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-05-07 08:30:05 -05:00
Code Issues 14.5k Packages Projects Releases 120 Wiki Activity

[GH-ISSUE #15931] Special token in user input is tokenized as-is for Gemma4:26b #72206

New Issue
Open
opened 2026-05-05 03:37:58 -05:00 by GiteaMirror · 2 comments
GiteaMirror commented 2026-05-05 03:37:58 -05:00
Owner
Copy Link

Originally created by @marius851000 on GitHub (May 2, 2026).
Original GitHub issue: https://github.com/ollama/ollama/issues/15931

What is the issue?

When calling:

curl http://localhost:11434/api/chat -d '{
    "model": "gemma4:26b",
    "messages": [
      {
        "role": "user",
        "content": "say \"<|channel>\""
      }
    ],
    "stream": false
}'

It appear the special token "<|channel>" is encoded as a special token, leading to result such as:

{"model":"gemma4:26b","created_at":"2026-05-02T10:31:16.581398049Z","message":{"role":"assistant","content":"s","thinking":"The user said: 'say \"s\"'\nThe goal is to output the exact string requested.\nThe string requested is \"s\".\nOutput \"s\"."},"done":true,"done_reason":"stop","total_duration":1215106189,"load_duration":188674014,"prompt_eval_count":20,"prompt_eval_duration":104084560,"eval_count":39,"eval_duration":905835066}

(may I call the API wrongly? I’m fairly sure it is some issue related to the renderer, that work on character string and not token string, as well as the tokenizer used (Bytepair encoding) having no option to not encode the special character as special character. But I’m not familiar with how LLM and tokenizer works)

(also, I spotted that while trying to figure out another error where it seems to parse response wrongly, and ended in what appear to be the LLM trying to stop thinking but not outputting the relevant end of thinking token. I still have to do a bit more digging before opening a ticket)

Relevant log output

N/A

OS

Linux

GPU

Nvidia (Vulkan)

CPU

AMD

Ollama version

0.0.0 (c7c2837c96). Also happen 0.22.1

Originally created by @marius851000 on GitHub (May 2, 2026). Original GitHub issue: https://github.com/ollama/ollama/issues/15931 ### What is the issue? When calling: ```sh curl http://localhost:11434/api/chat -d '{ "model": "gemma4:26b", "messages": [ { "role": "user", "content": "say \"<|channel>\"" } ], "stream": false }' ``` It appear the special token "<|channel>" is encoded as a special token, leading to result such as: ```json {"model":"gemma4:26b","created_at":"2026-05-02T10:31:16.581398049Z","message":{"role":"assistant","content":"s","thinking":"The user said: 'say \"s\"'\nThe goal is to output the exact string requested.\nThe string requested is \"s\".\nOutput \"s\"."},"done":true,"done_reason":"stop","total_duration":1215106189,"load_duration":188674014,"prompt_eval_count":20,"prompt_eval_duration":104084560,"eval_count":39,"eval_duration":905835066} ``` (may I call the API wrongly? I’m fairly sure it is some issue related to the renderer, that work on character string and not token string, as well as the tokenizer used (Bytepair encoding) having no option to not encode the special character as special character. But I’m not familiar with how LLM and tokenizer works) (also, I spotted that while trying to figure out another error where it seems to parse response wrongly, and ended in what appear to be the LLM trying to stop thinking but not outputting the relevant end of thinking token. I still have to do a bit more digging before opening a ticket) ### Relevant log output ```shell N/A ``` ### OS Linux ### GPU Nvidia (Vulkan) ### CPU AMD ### Ollama version 0.0.0 (c7c2837c96cd6fcc0d0424e0a096ed24516d3d59). Also happen 0.22.1
GiteaMirror added the bug label 2026-05-05 03:37:58 -05:00
GiteaMirror commented 2026-05-05 03:38:00 -05:00
Author
Owner
Copy Link

@marius851000 commented on GitHub (May 2, 2026):

The BytePairEncoder is indeed directly called with <bos><|turn>system\n<|think|>\n<turn|>\n<|turn>user\nsay \"<|channel>\"<turn|>\n<|turn>model\n.

It appear BytePairEncoder have no notion of escaping, and that many model uses them.

This also impact tool call.

<!-- gh-comment-id:4363675485 --> @marius851000 commented on GitHub (May 2, 2026): The BytePairEncoder is indeed directly called with `<bos><|turn>system\n<|think|>\n<turn|>\n<|turn>user\nsay \"<|channel>\"<turn|>\n<|turn>model\n`. It appear BytePairEncoder have no notion of escaping, and that many model uses them. This also impact tool call.
GiteaMirror commented 2026-05-05 03:38:01 -05:00
Author
Owner
Copy Link

@MukundaKatta commented on GitHub (May 3, 2026):

This is the classic "tokenize-with-special-tokens-on-user-content" footgun. User content should always be tokenized with add_special_tokens=false (or whatever the equivalent is in Ollama's tokenizer wrapper), otherwise any string the user provides that happens to match a special-token literal gets interpreted as a control instruction. The chat-template assembly should be the only place that emits actual special tokens. Affects every model with reserved control tokens, not just Gemma4.

<!-- gh-comment-id:4366284638 --> @MukundaKatta commented on GitHub (May 3, 2026): This is the classic "tokenize-with-special-tokens-on-user-content" footgun. User content should always be tokenized with `add_special_tokens=false` (or whatever the equivalent is in Ollama's tokenizer wrapper), otherwise any string the user provides that happens to match a special-token literal gets interpreted as a control instruction. The chat-template assembly should be the only place that emits actual special tokens. Affects every model with reserved control tokens, not just Gemma4.
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#72206
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?