SHACL Shape Dataset for Knowledge Graph Profiling

Description

This dataset contains 1,412 SHACL shapes in Turtle (.ttl) format, covering two knowledge graphs — DBpedia (2022, 20 classes) and YAGO (4.5,  37 classes) — across three categories:

Ground Truth — Manually curated shapes produced via SPARQL profiling, two-round crowd annotation, and expert consolidation. Released in two  DBpedia variants (dbpedia-aligned, dbpedia-complete) and one YAGO variant.

Statistical Baselines — Shapes from two fully automated systems: sheXer (https://github.com/DaniFdezAlvarez/shexer) (frequency-based  induction) and RDFShapeInduction (four ML classifiers: decision tree, gradient boosting, MLP, random forest).

LLM-Based Shapes — Shapes generated by large and small language models (GPT-4o mini, DeepSeek-V3, Qwen3, Gemma-3, Llama-3.2) using three  few-shot prompting strategies (local, global, triples) and a zero-shot agentic pipeline with skill injection, review/repair, and  consolidation steps. A deterministic heuristic baseline is also included.

This dataset accompanies the paper "Zero-Shot Agentic SHACL Shape Generation from Knowledge Graphs" and is intended for reproducibility and benchmarking of shape-induction methods.

Dataset Structure

```DATASET/├── GroundTruth/│   └── SHACL/│       ├── dbpedia-aligned/       # 20 shapes — DBpedia ground truth (aligned variant)│       ├── dbpedia-complete/      # 20 shapes — DBpedia ground truth (complete variant)│       └── yago/                  # 37 shapes — YAGO ground truth├── Baseline/│   ├── shexer/│   │   ├── dbpedia/shacl/         # 21 shapes — sheXer on DBpedia│   │   └── yago/shacl/            # 36 shapes — sheXer on YAGO│   └── rdfshapeinduction/│       ├── dbpedia/shacl/│       │   ├── (top-level)        # 19 shapes — best-model variant│       │   ├── decision_tree/     # 20 shapes│       │   ├── gradient_boosting/ # 20 shapes│       │   ├── mlp/               # 20 shapes│       │   └── random_forest/     # 20 shapes│       └── yago/shacl/│           ├── decision_tree/     # 35 shapes│           ├── gradient_boosting/ # 35 shapes│           ├── mlp/               # 35 shapes│           └── random_forest/     # 35 shapes└── LLMbased/    ├── local/{model}/{kg}/predicate_count/5/       # few-shot local mode    ├── global/{model}/{kg}/predicate_count/5/      # few-shot global mode    ├── triples/{model}/{kg}/predicate_count/5/     # few-shot triples mode    └── agent/{kg}/SHACL/        ├── global/{model}/                         # agentic, no consolidation        └── global_skills_review_conso/{model}/     # agentic + skills + review + consolidation```