swjtu-ml-experiment-skill

name	swjtu-ml-experiment-skill
description	Use when finalizing SWJTU machine-learning experiment reports, course-design reports, DOCX/PDF submissions, code appendices, grading-risk audits, citation checks, and zip/7z package verification.

SWJTU Machine Learning Experiment Skill

Overview

This skill is for strengthening SWJTU machine-learning experiment and course-design submissions before delivery. It focuses on requirement compliance, template fidelity, reproducible code, natural Chinese report writing, PDF rendering, and clean packaging.

Start Rules

• Read project-local instructions such as AGENT.md or AGENTS.md before editing or verifying coursework files.
• Treat teacher requirements, report templates, dataset rules, naming rules, and required file lists as hard constraints.
• If the task touches .docx or PDF layout, use a document/PDF workflow and visually inspect the rendered result.
• If a DOCX has been manually edited, do not regenerate it from scripts unless explicitly requested. Render the PDF from the current DOCX.
• Do not claim guaranteed grades or guaranteed AI-detector outcomes. Report only what was checked.

Workflow

1. Requirement Lock

   • Extract the assignment requirements: dataset, algorithms, metrics, figures, report sections, references, naming format, and submission file set.
   • Compare requirements against the actual report and code.
   • Fix missing hard requirements before polishing wording or layout.

2. Code Sanity

   • Keep submitted code readable and explainable by a student: direct flow, clear variable names, limited helper layers, and no unnecessary defensive framework.
   • Preserve true experiment logic and results. Do not change training logic only for style.
   • Self-implement coursework metrics unless the teacher or user explicitly allows packaged metric helpers.
   • Scan for risky helpers such as sklearn.metrics, accuracy_score, precision_score, recall_score, f1_score, and confusion_matrix( when self-implementation is required.

3. Report Polish

   • Preserve the teacher template: title, student information area, heading hierarchy, margins, tables, figures, and required chapters.
   • Make the prose concrete and natural: cite experiment settings, observed results, figures, tables, and error patterns.
   • Avoid generic AI-like filler, placeholders, exaggerated claims, and unsupported conclusions.
   • Ensure every listed reference has a real body citation when feasible.

4. Visual Render

   • Export DOCX to PDF with LibreOffice or an equivalent renderer.
   • Render PDF pages to images and inspect key pages: cover, tables, figures, code appendix start/end, and references.
   • For code appendices, prefer white background, readable monospaced font, left alignment, no line numbers unless requested, and no clipped text.

5. Package

   • Build a clean submission folder containing only the required deliverables, usually docx, pdf, and source code.
   • Remove Word lock/temp files such as ~$*.docx.
   • Create .7z when Windows ZIP compatibility is a concern; keep .zip as backup if useful.

6. Final Verification

   • Test archives with 7z t and unzip -t when available.
   • List and count files in the final package.
   • Hash-check copied PDF/code against the source files after recent edits.
   • Run or import submitted code enough to verify syntax, data paths, and key outputs when dependencies allow.
   • Search for residues such as ChatGPT, AI生成, TODO, 占位, 这里填写, stale function names, forbidden metric helpers, and temp files.

Final Response

Keep the final response concise. Include the artifact path or repository URL, checks that passed, and any honest residual risks.