Datasheet — data accompanying the TempGraphRAG methods paper

Anonymous double-blind artifact. Covers the data shipped to reproduce the methods paper. The benchmark itself is described in concurrent anonymous work (anon-bench); here it is a read-only input.

Provenance

Source KG: tkgl-smallpedia from TGB 2.0, derived from Wikidata (public, CC0). 550,376 timestamped edges, 47,433 entities, 283 relations, span 1900–2024.
Benchmark questions (benchmark_labelled.jsonl): 8,710 multi-hop temporal-KGQA questions, each with a gold supporting subgraph ($S^{*}$), same-$(s,r)$ distractors ($S_{\mathrm{dist}}$), and stale-fact set ($S_{\mathrm{stale}}$), stratified over a 4×3 operator × complexity matrix. 70/10/20 train/dev/test split (6,096 / 871 / 1,743). Generated from a TimelineKGQA-derived pipeline (see anon-bench).
SFT corpora (sft_data_5k_cot.jsonl, *_terse_lever.jsonl, *_multitq_5k.jsonl): retrieval-conditioned traces derived from the training split; used to fine-tune the generator.
Evaluation outputs (outputs/eval/*.json): per-question prediction + gold for every reported run (3-seed headline, ablations, judge runs, MultiTQ).

Composition / splits

Test results in the paper are on the 1,743-question test split. The train/test splits share the source KG; the paper measures and reports contamination (0.40% verbatim-question overlap, 5.1% anchor overlap) and treats absolute TempBench EM as an in-distribution upper bound, with MultiTQ as the cross-distribution check.

Intended use

Research reproduction of retrieval-augmented temporal multi-hop KGQA. Not intended as a production correctness oracle. The triple-level grounding metric (appendix) is a population-level diagnostic, not a per-instance guarantee.

Personal / sensitive data

Questions concern historical and public-figure facts as exposed by Wikidata; no personal data beyond Wikidata's public content. No human-subject data in this archive. (Human inter-annotator validation is reported with anon-bench, not here.)

License

CC0-derived (Wikidata) via TGB 2.0; redistribution under permissive terms consistent with TGB 2.0. Trained LoRA adapters are released for research use.