curate_dataset.md

RepoQA Dataset Curation

Search Needle Functions

Step 1: Cherry-pick repositories

See scripts/cherrypick/README.md for more information.

Tip

Output: Extend scripts/cherrypick/lists.json for a programming language.

Step 2: Extract repo content

python scripts/curate/dataset_ensemble_clone.py

Tip

Output: repoqa-{datetime}.json by adding a "content" field (path to content) for each repo.

Step 3: Dependency analysis

Check scripts/curate/dep_analysis for more information.

python scripts/curate/dep_analysis/{language}.py  # python

Tip

Output: {language}.json (e.g., python.json) with a list of items of {"repo": ..., "commit_sha": ..., "dependency": ...} field where the dependency is a map of path to imported paths.

Note

The {language}.json should be uploaded as a release.

To fetch the release, go to scripts/curate/dep_analysis/data and run gh release download dependency --pattern "*.json" --clobber.

Step 4: Merge step 2 and step 3

python scripts/curate/merge_dep.py --dataset-path repoqa-{datetime}.json

Tip

Input: Download dependency files in to scripts/curate/dep_analysis/data.

Output: Update repoqa-{datetime}.json by adding a "dependency" field for each repository.

Step 5: Function collection with TreeSitter

# collect functions (in-place)
python scripts/curate/function_analysis.py --dataset-path repoqa-{datetime}.json
# select needles (in-place)
python scripts/curate/needle_selection.py --dataset-path repoqa-{datetime}.json

Tip

Output: --dataset-path (in-place) by adding a "functions" field (path to a list function information) for each repo.

Step 6: Annotate each function with description to make a final dataset

python scripts/curate/needle_annotation.py --dataset-path repoqa-{datetime}.json

Tip

You need to set OPENAI_API_KEY in the environment variable to run GPT-4. But you can enable --use-batch-api to save some costs.

Output: --output-desc-path is a seperate json file specifying the function annotations with its sources.

Step 7: Merge needle description to the final dataset

python scripts/curate/merge_annotation.py --dataset-path repoqa-{datetime}.json --annotation-path {output-desc-path}.jsonl

Tip

Output: --dataset-path (in-place) by adding a "description" field for each needle function.