## Text-Based Template: RRM Prompt Template for AI Response Evaluation
### Overview
The image depicts a structured template for evaluating and comparing responses from two AI assistants. It provides explicit instructions, evaluation criteria, and formatting rules for a "Response Ranking Model" (RRM) task. The template emphasizes objectivity, safety, and precision in assessing AI-generated outputs.
### Components/Axes
- **Header**:
- Title: "RRM Prompt Template" (bold, centered, dark blue background).
- Subtitle: "You are a helpful assistant in evaluating the quality of the responses for a given instruction."
- **Main Body**:
- **Instructions**:
- Goal: Select the better response (Assistant 1 or Assistant 2) for a given instruction.
- Rules:
1. Prioritize harmlessness/safety if the instruction contains harmful content.
2. Evaluate helpfulness, accuracy, detail, and precision if the instruction is safe.
3. Responses must not exceed the instruction’s requirements.
4. Avoid bias; responses are equally likely to be better regardless of order or length.
- **Bias Sources**:
- Response order, length, and presentation timing.
- **Output Format**:
- Only output `\boxed{Assistant 1}` or `\boxed{Assistant 2}` based on evaluation.
- **Placeholders**:
- `## Query` (input instruction).
- `### Assistant responses` (two responses labeled `### Assistant 1` and `### Assistant 2`).
- `### Analysis` (step-by-step reasoning section).
### Detailed Analysis
- **Textual Content**:
- The template enforces strict evaluation criteria, such as:
- Harmful content prioritization (Rule 1).
- Precision in response length (Rule 3).
- Objectivity in bias avoidance (Rule 4).
- Placeholders use hierarchical headings (`##`, `###`) for structured input.
- Output is restricted to a single boxed assistant identifier.
- **Formatting**:
- Dark blue header with white text.
- Body text in black on a light gray background.
- Placeholders use bold labels (e.g., `## Query`).
### Key Observations
- No numerical data, charts, or diagrams are present.
- The template is purely textual, focusing on procedural guidelines.
- Emphasis on safety and objectivity aligns with ethical AI evaluation practices.
### Interpretation
This template standardizes the evaluation of AI responses by:
1. **Defining Clear Priorities**: Safety first, then accuracy/helpfulness.
2. **Mitigating Bias**: Explicitly addressing response order and length as potential confounders.
3. **Enforcing Precision**: Responses must match the instruction’s scope.
4. **Structured Output**: The `\boxed{}` format ensures unambiguous results.
The absence of numerical data suggests this is a procedural framework rather than an analytical tool. Its design reflects a focus on reproducibility and fairness in AI assessment, critical for red-teaming or quality assurance workflows.
