File size: 5,598 Bytes
cf1cbf1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 | ---
license: cc-by-sa-4.0
language:
- en
pretty_name: PIPE-RDF Benchmarks
size_categories:
- 1K<n<10K
task_categories:
- question-answering
- text-generation
tags:
- text-to-sparql
- sparql
- rdf
- knowledge-graphs
- kgqa
- benchmark
configs:
- config_name: schema_c
data_files:
- split: test
path: benchmark/schema_c/benchmark_phase3_balanced.jsonl
- config_name: spb_full
data_files:
- split: test
path: benchmark/spb_full/benchmark_phase3_balanced.jsonl
---
# PIPE-RDF Benchmarks
This dataset holds the benchmark and evaluation artifacts released with the paper
**PIPE-RDF: Execution-Grounded Generation of Schema-Specific NL–SPARQL Benchmarks**
(Findings of EACL 2027).
- Code: https://github.com/suraj-ranganath/PIPE-RDF
- Dataset: https://huggingface.co/datasets/suraj-ranganath/PIPE-RDF-benchmarks
PIPE-RDF builds natural-language/SPARQL benchmarks for a target RDF graph. It
fills question templates only with bindings returned by reverse queries on the
graph, generates candidates with category-aware retrieval, and keeps a pair only
if its SPARQL query parses, executes, returns a non-empty answer of the expected
type, and matches the expected form for its category.
## Main benchmarks
| Config | Graph | Pairs | Categories |
|---|---|---|---|
| `schema_c` | Company–location slice (5,000 companies; DBpedia-style IRIs) | 1,800 | 9 × 200 |
| `spb_full` | LDBC Semantic Publishing Benchmark graph (~25M triples) | 1,800 | 9 × 200 |
The nine categories are `generic`, `counting`, `comparative`, `superlative`,
`ordinal`, `multi-hop`, `intersection`, `difference`, and `yesno`. Both
benchmarks were generated with Qwen3.5-4B. In the released files, none of the
3,600 pairs has a parse failure, an execution failure, or an empty answer.
```python
from datasets import load_dataset
schema_c = load_dataset("suraj-ranganath/PIPE-RDF-benchmarks", "schema_c", split="test")
spb_full = load_dataset("suraj-ranganath/PIPE-RDF-benchmarks", "spb_full", split="test")
```
### Record fields
| Field | Description |
|---|---|
| `category` | Query category (one of the nine above) |
| `question` | Natural-language question |
| `sparql` | SPARQL query, with prefix declarations |
| `answers` | Answers returned by the endpoint, as strings (`"True"`/`"False"` for `ASK`) |
| `answer_count` | Number of returned answers |
| `parse_valid`, `exec_success` | Validation flags (all `true` in the released benchmarks) |
| `error`, `error_type` | Validation error, if any (all `null` in the released benchmarks) |
| `repair_attempts` | Number of LLM repair calls used for the record |
| `retrieved_examples` | Retrieved examples (question, SPARQL, category, score) used in the generation prompt |
| `ast_node_count`, `ast_max_depth` | Structural size of the parsed query |
| `llm_latency_ms`, `sparql_exec_ms`, `question_latency_ms` | Latency diagnostics |
| `prompt_chars`, `prompt_tokens_est` | Prompt size |
| `result_hash` | Hash of the returned result set |
Every benchmark directory also has a CSV copy of the JSONL and a
`benchmark_phase3_balanced_summary.json`. The summary reports category counts,
parse/execution/empty failures, repairs, the `ASK` answer distribution,
predicate inventory, and latency totals.
## Other artifacts
| Path | Contents |
|---|---|
| `schema_profiles/{schema_c,spb_full}/` | Predicate counts, type counts, class–predicate matrices, and schema summaries |
| `semantic_audit/` | The 216-record stratified audit packet, raw outputs from both LLM judges (GPT-5-mini, Grok-4.20-0309 non-reasoning), adjudicated labels, and summary |
| `cross_model/` | Strict 50-per-category probe benchmarks from Qwen3.5-2B, 4B, and 9B on both schemas |
| `downstream_utility/` | Zero-shot and category-RAG NL-to-SPARQL results (per-example outputs and summaries) for Qwen3.5-2B and 4B on both schemas |
| `diagnostics/` | Operational-efficiency summaries, natural-language and structural diagnostics, and per-category paired deltas for the downstream evaluation |
## Intended use
Use these files to evaluate NL-to-SPARQL systems on the two schemas, to supply
schema- and category-matched retrieval examples, or to reproduce the tables in
the paper. To run the queries, load the matching graph into a SPARQL endpoint
as described in the code repository. To get a benchmark for a different RDF
graph, run the pipeline on that graph.
## Limitations
- Both graphs are public. The release says nothing about private or proprietary
graphs.
- The taxonomy does not target property paths, named graphs, language-tag
filters, broad `UNION` use, or subquery-centered queries.
- Questions and labels are English only.
- Semantic correctness was audited by two LLM judges on a 216-pair sample, not
by human annotation of every pair. Both judges passed 212 of the 216 sampled
pairs.
## Source data and license
The questions, queries, and diagnostics were generated by the PIPE-RDF pipeline.
Entity IRIs, labels, and answers come from public RDF data, which includes
DBpedia-style resources and the LDBC Semantic Publishing Benchmark reference
data. The release is under CC BY-SA 4.0 so that it stays compatible with the
share-alike terms of DBpedia-derived content. Follow the upstream data licenses
when you redistribute the underlying graphs.
## Citation
```bibtex
@inproceedings{ranganath2027piperdf,
title = {{PIPE-RDF}: Execution-Grounded Generation of Schema-Specific {NL--SPARQL} Benchmarks},
author = {Ranganath, Suraj},
booktitle = {Findings of the Association for Computational Linguistics: EACL 2027},
year = {2027}
}
```
|