Documentation-Grounded Code Question Answering Dataset

This dataset contains aligned code–documentation pairs collected from a curated set of open-source repositories, along with an augmented set of samples from CodeSearchNet (Python and Java). The pipeline includes repository filtering, documentation entity extraction (docstrings, comments, Javadoc), and alignment of code with its associated documentation. To improve quality (and for downstream training), a subset of the data is validated using an LLM-as-a-judge approach, filtering out low-confidence alignments. The final dataset includes generated natural-language questions for each code–documentation pair, designed to reflect realistic developer queries that require reasoning over both code and documentation. The release includes: Repository metadata and source files Extracted documentation entities Aligned code–documentation pairs LLM-verified subsets Augmented samples from CodeSearchNet Generated question datasets Train/validation/test splits with repository-level separation to prevent data leakage This dataset is intended for research on code understanding, documentation-grounded reasoning, and code question answering. File index: accepted_repos.jsonl : list of accepted repositories accepted_repo_files.jsonl: their corresponding files (so we can perform alignment later on) file_entities.jsonl : documentation entities from each file (inline comment, docstring, javadoc, commit log) final_aligned_dataset.jsonl: all aligned code-documentation pairs csn_samples.jsonl: 2k pairs retrieved from CodeSearchNet training_dataset_final: subset of our aligned pairs (~8k) and CodeSearchNet pairs training_dataset_verified: subset of pairs following LLM judgement (poorly aligned pairs removed) pilot_judgements.jsonl: trace of the LLM judge's decisions results.jsonl: (c,d,q,a) tuples following synthetic question generation train.jsonl: training dataset val.jsonl: validation dataset test.jsonl: test dataset (for evaluation)