evaluating-multi-step-edits

星标14

分支1

更新时间2026年5月14日 17:35

Guidelines for evaluating prompts with multiple distinct instructions.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

TIGER-AI-Lab

TIGER-AI-Lab/RewardHarness

下载

Evaluating Multi-Step Edits

When a prompt contains multiple instructions (e.g., "Remove the object on the left AND increase the brightness of the central figure"):

Deconstruct the Prompt: Break the prompt down into its individual components.
Check Each Component: Evaluate whether each edited image successfully completes every component.
Partial Success: If an image only completes one of the two instructions, it should receive a lower Instruction Following score (e.g., 2 or 3) compared to an image that completes both.
Over-editing: Ensure that global adjustments (like brightness) are only applied to the specified areas (e.g., "central figure") and not the entire image, unless requested.

同仓库更多 Skills

同仓库

Score completeness of instruction execution. Partial edits get partial credit; missed sub-steps cost points.

Guidance on penalizing artifacts while allowing conceptual unrealism if requested by the prompt.

Extracts and analyzes text within images to verify spelling, placement, and clarity.

Guidelines for evaluating background replacements and scene changes.

Guidelines for evaluating object replacements and modifications.

Guidelines for evaluating artistic style transfers.

description	Guidelines for evaluating prompts with multiple distinct instructions.
name	evaluating-multi-step-edits
type	skill

When a prompt contains multiple instructions (e.g., "Remove the object on the left AND increase the brightness of the central figure"):

Deconstruct the Prompt: Break the prompt down into its individual components.
Check Each Component: Evaluate whether each edited image successfully completes every component.
Partial Success: If an image only completes one of the two instructions, it should receive a lower Instruction Following score (e.g., 2 or 3) compared to an image that completes both.
Over-editing: Ensure that global adjustments (like brightness) are only applied to the specified areas (e.g., "central figure") and not the entire image, unless requested.