Reasoning / thinking control

Purpose

Control how reasoning (<think>...</think>) is kept, removed, and trained. This page covers think_mode, think_loss, and think_max_tokens behavior.

When to use

• You want answer-only behavior for agent/tool datasets.
• You want to preserve full reasoning traces for reasoning datasets.
• You need to cap reasoning token length for memory/perf constraints.

Syntax (config fields)

think_mode: keep                 # keep | drop
think_loss: answer_only          # all | answer_only | answer_plus_think
think_max_tokens: 0              # 0 = no cap

Options

think_mode

Value	Effect
keep	Keep reasoning content (with optional cap)
drop	Remove reasoning content entirely

think_loss

Value	Gradient scope
all	Full assistant span: think + answer
answer_only	Tokens after </think> only
answer_plus_think	Think content + answer, excluding literal tags

think_max_tokens

• 0 means no cap.
• Positive values cap reasoning token count per message.

Decision flow

1. Decide content retention (`think_mode`):
   - `keep` to preserve reasoning content.
   - `drop` for answer-only datasets/behavior.
2. Decide gradient scope (`think_loss`):
   - `all`, `answer_only`, or `answer_plus_think`.
3. Decide cap (`think_max_tokens`):
   - `0` for unlimited.
   - positive value to constrain long reasoning traces.
4. Validate against eval target:
   - train/eval behavior should match (answer-only vs full reasoning).

Examples

# Agent/tool training: keep think text, train only final answer
think_mode: keep
think_loss: answer_only

# Full reasoning training
think_mode: keep
think_loss: all

# Pure answer-style behavior
think_mode: drop
think_loss: all

# Reasoning cap example
think_max_tokens: 512

Edge cases / limitations

WARNING

Aggressive think_max_tokens can truncate reasoning mid-thought and teach abrupt stop patterns. Prefer cleaner datasets over heavy truncation when possible.

• think_loss has no practical effect on reasoning text when think_mode: drop.
• Keep reasoning settings aligned with your evaluation target (answer-only vs full reasoning).

Related

• Config reference
• Dataset pipeline
• Quickstart