## Technical Document: Prompt Template for GPT-4o Training Dataset Construction ### Overview The image displays a structured text document outlining a prompt template designed for an AI model (specifically GPT-4o) to construct a training dataset. The document is formatted as a technical specification, defining a task for evaluating and correcting multi-step solutions to multimodal mathematical problems. The text is entirely in English. ### Content Details The document is organized into the following sections, transcribed verbatim: **Header:** `Prompt for GPT-4o to construct training dataset:` Followed by a dashed line separator. **Introduction:** `You are an expert in solving multimodal mathematical problems. You will be given:` `1. A multimodal mathematical problem and its corresponding image.` `2. A multiple-step solution (each step on a new line).` **Task Section (Labeled in Red):** `**Task**:` (The label "**Task**:" is in red text) `The tasks you need to do are:` `1. Analyze the purpose of each step and what specific actions were taken in each step.` `2. Analyze each step's correctness in terms of image alignment and reasoning logic.` `- Image alignment: Whether the information and reasoning used in the step are consistent with the content of the provided image.` `- Reasoning logic: Whether the reasoning is logically sound, calculations are correct, and information used matches that from previous steps and question.` `When outputting judgements, you must choose one output from "Correct" or "Incorrect".` `3. For the first incorrect step, correct it based on your analysis of its error, and output the corrected step at the end of your output.` **Question Section (Labeled in Red):** `**Question**:` (The label "**Question**:" is in red text) `The multimodal mathematical problem is as follows:` `<Question>` (This is a blue placeholder tag) **Solution Steps Section (Labeled in Red):** `**Solution Steps**:` (The label "**Solution Steps**:" is in red text) `The multiple-step solution is as follows:` `<Solution Steps>` (This is a blue placeholder tag) **Output Format Section (Labeled in Red):** `**Output Format**:` (The label "**Output Format**:" is in red text) `You must output your content in the following format:` `### Step 1 ###` `Step intent analysis:[Describe what the step aims to do and the specific actions]` `Image alignment analysis:[Analyze the consistency of image alignment]` `Judgement of image alignment:[Correct/Incorrect]` `Reasoning logic analysis:[Analyze the rationality of logic, correctness of calculations and consistency with prior step]` `Judgement of reasoning logic:[Correct/Incorrect]` `Final judgement of the current step:[Correct/Incorrect]` `### Step 2 ###` `...` `Corrected step of the first incorrect step in solution:` `Step n:[assume that the first incorrect step is step n, and fill in the corrected step n in the square bracket]` ### Key Observations 1. **Document Type:** This is a meta-prompt or a system prompt template, not a chart, diagram, or data visualization. It contains instructions and a structured format for another AI to follow. 2. **Formatting Cues:** Key section labels (`**Task**:`, `**Question**:`, `**Solution Steps**:`, `**Output Format**:`) are highlighted in red. Placeholders for dynamic content (`<Question>`, `<Solution Steps>`) are highlighted in blue. 3. **Structured Output:** The required output format is highly structured, mandating a step-by-step analysis with specific fields for intent, image alignment, reasoning logic, and final judgment for each step in a solution. 4. **Error Correction Protocol:** The template includes a specific instruction to identify and correct the *first* incorrect step found in the provided solution. ### Interpretation This document defines a rigorous evaluation protocol for multimodal mathematical reasoning. Its purpose is to generate high-quality training data by having an AI model act as a "grader" or "tutor." The process involves: * **Deconstruction:** Breaking down a solution into discrete steps. * **Multi-faceted Evaluation:** Assessing each step on two independent axes: fidelity to the provided image (visual grounding) and internal logical/mathematical soundness. * **Structured Judgement:** Enforcing a binary (Correct/Incorrect) output for clarity and consistency in the generated dataset. * **Pedagogical Correction:** Requiring the generation of a corrected step, which adds a constructive, teaching-oriented element to the dataset. The template is designed to produce data that can train models not just to solve problems, but to *explain* and *critique* solution processes, emphasizing the importance of aligning reasoning with visual evidence. The strict output format ensures the generated data is machine-readable and consistent, suitable for fine-tuning or evaluation benchmarks.