GRISP material

General

• Demo of GRISP on Wikidata (model trained on WDQL)
• Evaluation app for evaluating and comparing models on benchmarks
• Benchmark ground truths and model outputs (GRISP, its ablations, and the baselines)
• Benchmarks with the suffix -qlever have gold queries rewritten from Wikidata Query Service (Blazegraph) syntax to standard SPARQL so they run on QLever; the original query is kept in info.original_sparql and every change is listed in rewrite_report.json. SPINACH: predeclared prefixes, label service, and named subqueries. WDQL: predeclared prefixes. LC-QuAD 2.0: contains(YEAR(?x), ...) becomes contains(str(YEAR(?x)), ...), in both the test and the train set

Knowledge graph indices

• Wikidata, used by all Wikidata checkpoints below
• Freebase, used by all Freebase checkpoints below
• These differ from the GRASP indices; extract them into their own GRASP_INDEX_DIR. Their info.json points to the public QLever endpoints, while our experiments used our own QLever instance with a Wikidata dump from February 2026 and a Virtuoso instance for Freebase

Training data

• Wikidata and Freebase: GRISP training data of all checkpoints below, one folder per benchmark. train.jsonl / val.jsonl are the converted samples (output of grisp.data, val.jsonl only where the benchmark has its own validation set), train.materialized.jsonl / val.materialized.jsonl are the materialized samples the models were trained and validated on (output of grisp.materialize; without val.jsonl, 10% of the training samples were split off for validation)

Model checkpoints (Freebase)

• Qwen2.5 (7B), trained on CWQ
• Qwen2.5 (7B), trained on WebQSP

Model checkpoints (Wikidata)

• Qwen2.5 (7B), trained on WWQ, used for WWQ and WWQ_dev
• Qwen2.5 (7B), trained on QALD-7
• Qwen2.5 (7B), trained on QALD-10
• Qwen2.5 (7B), trained on LC-QuAD 2.0 (gold queries rewritten for QLever, see below)
• Qwen2.5 (7B), trained on SimpleQuestions
• Qwen2.5 (7B), trained on WDQL, used for SPINACH, QAWiki, WDQL, and the demo

Model checkpoints for the analyses (Wikidata, trained on WWQ)

• Qwen2.5 (0.5B), model size study
• Qwen2.5 (1.5B), model size study
• Qwen2.5 (3B), model size study
• Qwen2.5 (7B), skeleton generation only, training setup study
• Qwen2.5 (7B), IRI selection only, training setup study

Model checkpoints for the analyses (Wikidata, trained on QALD-10)

• Qwen2.5 (0.5B), model size study
• Qwen2.5 (1.5B), model size study
• Qwen2.5 (3B), model size study