Skip to content

Deployed adapter inference drops opening <think> tag #687

@austinmw

Description

@austinmw

A deployed adapter response includes a closing </think> marker in message.content, but the corresponding opening <think> marker is missing.

Model:

Qwen/Qwen3.5-2B:clk03rc5bmfzlior7p9070z9

Repro:

prime inference chat "Qwen/Qwen3.5-2B:clk03rc5bmfzlior7p9070z9" \
  "Sort these names in alphabetical order by FIRST name: MasatoshiKokubu, MasaakiUmehara, WayneRossman

Use exactly this format:
<alphabetical_sorted>
Name1
Name2
Name3
</alphabetical_sorted>" \
  --max-tokens 512

CLI / inference output:

MasatoshiKokubu
MasaakiUmehara
WayneRossman
</think>

<alphabetical_sorted>
MasaakiUmehara
MasatoshiKokubu
WayneRossman
</alphabetical_sorted>

Training UI shows the corresponding output with the opening tag present:

<think>MasatoshiKokubu
MasaakiUmehara
WayneRossman</think>

<alphabetical_sorted>
MasaakiUmehara
MasatoshiKokubu
WayneRossman
</alphabetical_sorted>

Expected: inference should either preserve the complete <think>...</think> block or strip/hide the thinking block consistently. The issue is specifically that the opening <think> marker appears to be dropped while </think> remains.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions