Open Models Almanac
Methodspost-training

single mention · post-training

Reasoning-effort labeling

A control that lets a model answer directly (reasoning effort none) or reason at low, medium, or high effort, which the chat template requests through a fixed sentence in the system prompt.1 In SFT, reasoning-effort labels are assigned from dataset-specific reasoning length statistics and judged difficulty, while RL trains each task across all reasoning levels with different length penalties.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. Kolibri can answer directly (reasoning effort none) or reason at low, medium or high reasoning effort, which the chat template requests through a fixed sentence in the system prompt — Aleph-Alpha/Kolibri-1 · Hugging Face
  2. In SFT, we assigned reasoning effort labels based on observed dataset-specific reasoning length statistics and judged difficulty, while on RL side we trained each task across all reasoning levels with different length penalties — Aleph-Alpha/Kolibri-1 · Hugging Face

Also called Configurable reasoning-effort levels, Reasoning Effort.

Filed under post-training