OpenBMB / OpenBMB/MiniCPM

[Bug]: MiniCPM5 chat template may lose text after <tool_sep> due to loop-local assignments

Open
#379 0 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Jupyter Notebook
Stars
11.1k
Forks
763
Avg merge
6h 3m
Merged PRs (30d)
5

Description

Is there an existing issue?
  • I searched existing issues for tool_sep, processed_content, chat_template, and jinja and did not find this problem reported.
Describe the bug

MiniCPM5-2B's chat template appears to use loop-local set assignments to accumulate processed_content for use after the loops. Jinja2 does not propagate these assignments to the enclosing scope.

With assistant content BEFORE<tool_sep>AFTER and one tool call, the rendered prompt loses AFTER. The tool XML is emitted once by the later direct-output branch.

Template: openbmb/MiniCPM5-2B/chat_template.jinja, revision a063f08de1bd09dfc9ae4cf3da35e6064949e533.

To reproduce

Download the unchanged template as chat_template.jinja, then run this script in the same directory. It only requires Jinja2; no model weights or inference are involved.

from pathlib import Path

from jinja2 import Environment


template = Environment(trim_blocks=True, lstrip_blocks=True).from_string(
    Path("chat_template.jinja").read_text(encoding="utf-8")
)

for content in ["BEFORE", "BEFORE<tool_sep>AFTER"]:
    output = template.render(
        bos_token="",
        tools=[],
        messages=[
            {"role": "user", "content": "Question"},
            {
                "role": "assistant",
                "content": content,
                "tool_calls": [
                    {
                        "function": {
                            "name": "lookup",
                            "arguments": {"q": "weather"},
                        }
                    }
                ],
            },
        ],
        add_generation_prompt=False,
    )
    print("Input:", content)
    print("Contains AFTER:", "AFTER" in output)
    print("Tool XML count:", output.count('<function name="'))
    print(repr(output))

Observed output:

Input: BEFORE
Contains AFTER: False
Tool XML count: 1
'<|im_start|>user\nQuestion<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\nBEFORE\n<function name="lookup"><param name="q">weather</param></function><|im_end|>\n'
Input: BEFORE<tool_sep>AFTER
Contains AFTER: False
Tool XML count: 1
'<|im_start|>user\nQuestion<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\nBEFORE\n<function name="lookup"><param name="q">weather</param></function><|im_end|>\n'
Expected behavior

Assuming <tool_sep> marks an in-place tool-call position, I would expect both surrounding text segments to survive, with the tool XML inserted between BEFORE and AFTER. Could you confirm whether this is the intended behavior?

Environment
  • Python 3.12.2
  • Jinja2 3.1.3
  • Direct rendering of the unchanged template with Environment(trim_blocks=True, lstrip_blocks=True)
Additional context

The assignments at template lines 84, 86, and 112 are inside nested loops. When line 116 reads processed_content, its enclosing value remains content_parts[0]. Jinja documents this assignment-scoping behavior and the use of namespace objects for cross-scope state.

The final tool-output condition at line 127 also reads has_tool_sep, which is not defined anywhere in this template or supplied by this reproducer. The condition is therefore true for the tool-calling messages above. If accumulation is changed to persist across loops, that final output condition should be reviewed together with it so that each tool call is emitted once.

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by rendering the referenced chat_template.jinja with the provided Jinja2 reproduction and inspect lines 84, 86, 112, 116, and 127. Verify how loop-local assignments affect processed_content and has_tool_sep. Done means both BEFORE and AFTER survive around one emitted tool XML block, with the intended behavior confirmed.

Written by the indexing model from the issue text.

Assessment

Tech stack
python
Domain
ai
Issue type
Bug
Difficulty
3/5
Estimated time
1-2 days
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
62/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.