Dataset: Why Are AI-Generated Pull Requests Discarded? An Analysis Using the AIDEv Dataset

dataset.json The primary dataset containing complete metadata for all 4,314 Pull Requests analyzed in this study. Root object containing two main properties: Key Fields per PR Object: id (Integer): Unique GitHub PR identifier. number (Integer): PR number within the repository. title (String): PR title/heading. body (String): PR description and detailed content. agent (String): Tool/agent that created the PR (e.g., "Claude_Code", "Devin", "GitHub Copilot"). user_id (Integer): GitHub user ID of the PR author. user (String): GitHub username of the PR author. state (String): PR status (always "closed" in this dataset). created_at (ISO 8601 Datetime): PR creation timestamp. closed_at (ISO 8601 Datetime): PR closure timestamp. merged_at (null): Merge timestamp (null = not merged). repo_id (Float): GitHub repository ID. repo_url (String): API URL endpoint of the repository. html_url (String): Web URL of the PR on GitHub. comments (Array) Array of comment objects from PR discussion. Comments Array Structure: id (Integer): Unique comment ID. pr_id (Integer): Parent PR ID for linking. user (String): Comment author username. user_id (Integer): Comment author GitHub user ID. user_type (String): User type identifier ("User", "Bot", "Organization"). created_at (ISO 8601 Datetime): Comment creation timestamp. body (String): Full comment text content. summarize_gemini.json AI-generated semantic summaries of PR closure reasons using Google's Gemini LLM. This file maps each PR ID to a structured analysis explaining why the PR was closed without merging. Structure: pr_id (Integer): Unique GitHub PR identifier (matches dataset.json). html_url (String):Web URL of the PR on GitHub. original_comment (Array[String]): Original PR comment threads (context for analysis). analysis (String): AI-generated summary of closure reason. cluster_bert_generated.json Initial unsupervised clustering results generated by combining BERT embeddings and HDBSCAN algorithm. This file represents the raw algorithmic output before manual curation and consolidation. Root Level Metadata: source_run (Integer): Identifier for the clustering run. number_of_clusters (Integer): Total clusters generated by HDBSCAN. total_analyses (Integer): Total PR analysis texts processed. clustered_analyses (Integer): Texts successfully assigned to clusters. outliers_noise (Integer): Texts classified as outliers/noise. clusters (Array): Array of cluster objects with detailed content. Cluster Object Structure: cluster_id (String): Numeric identifier (0-169) or "OUTLIERS/NOISE". items (Integer): Number of PRs assigned to this cluster. keywords (Array[String]): Top 5-10 keywords extracted from cluster texts. texts (Array[String]): Representative analysis text samples from cluster. clusters_final.json Manually curated and consolidated clustering results derived from cluster_bert_generated.json. Domain experts reviewed the 170 initial clusters and merged semantically related patterns into 21 comprehensive categories representing common PR closure reasons. Root Level Metadata: source_run (Integer): Identifier for the clustering run. outliers_removed (Boolean): Outliers explicitly handled in consolidation. categories (Array): Consolidated category objects. Category Object Structure: new_cluster_id (Integer): Final category ID (1-21). label (String): Human-readable category name. description (String): Detailed description of PR closure reason pattern. original_clusters_ids (Array[Integer]): IDs of original clusters merged into this category. items (Integer): Total PRs assigned to this category. text (Array[String]): Representative closure reason texts from PRs. metadata (Object): Frequency distribution of unique analysis texts.