| name | repo-release-curation |
| description | Curate and restructure a near-publication ML repository for clean release, reproducible runs, and minimal layout while preserving eval/data and paper assets. |
Repo Release Curation
Use this skill when asked to tidy, restructure, or package a near-publication repository for open-source release while preserving reproducibility.
Workflow
-
Confirm preservation scope
- Always keep:
acl26_instructDiff/ (paper source), eval/, data/.
- Confirm any other keep/delete constraints before removing files.
-
Minimize root layout
- Target root:
acl26_instructDiff/, configs/, data/, eval/, src/, README.md, requirements.txt, setup.py, pyproject.toml, .gitignore.
- Move all code into
src/ and adjust imports accordingly.
-
Consolidate tools
- Merge plotting/analysis into
src/<tool_pkg>/analysis/.
- If utility helpers exist, merge into package
__init__.py or a single module to reduce file count.
-
Packaging
- Add
requirements.txt, setup.py, and pyproject.toml.
- Provide
console_scripts entry for CLI (instdiff).
- Ensure
pip install -e . works without extra steps.
-
Docs (bilingual)
- Write a single
README.md with Chinese + English sections.
- Include pipeline overview, quick start, config usage, and actual experiment results.
- Reference paper figures using local assets.
-
Configs and outputs
- Keep configs under
configs/<domain>/.
- Default outputs go to
runs/ and are ignored via .gitignore.
-
AI trace removal
- Remove assistant-specific logs or meta files.
- Remove temporary logs or internal notes not needed for users.
Checks
pip install -e . installs instdiff CLI.
README.md is the only root doc, bilingual, and includes results.
- Root contains only required top-level folders.
eval/ and data/ are preserved untouched.