# TempBench dataset card — v1.0.1 addendum Ready-to-paste section for the TempBench dataset card (DOI `10.57967/hf/10071`). Append under a "Known issues and evaluation protocol (v1.0.1)" heading. All numbers below were computed directly from the released artifacts (`benchmark_labelled.jsonl` and the reference-baseline evaluation outputs); regeneration commands are listed at the end. --- ## Known issues and evaluation protocol (v1.0.1) ### 1. Interval-operator answer leakage through `t_query` **Mechanism.** Every `interval` question's answer is a year, and the construction pipeline sets the question's `t_query` field exactly equal to that gold answer year. This holds for 864/864 (100.0%) of interval questions and for 0.0% of every other operator. Any evaluation protocol that exposes `t_query` to the system — in particular, prompt templates that render a `` prefix — makes the interval slice answerable by copying the timestamp out of the prompt, with no retrieval and no reasoning. **Affected fraction (exact, per split).** | Split | questions | interval | share | |---|---:|---:|---:| | train | 6,096 | 601 | 9.86% | | dev | 871 | 100 | 11.48% | | test | 1,743 | 163 | 9.35% | | **all** | **8,710** | **864** | **9.92%** | **Evidence that the leak is exploited, not merely exploitable.** Reference systems with no retrieval at all (prompt is ` {question}\nAnswer:`) score at or near 1.0 on the interval slice: a no-retrieval RL-tuned policy scores interval EM 1.000 against 0.364 overall, and a no-retrieval supervised baseline 0.994 against 0.305. On retrieval-conditioned reference systems, excluding the interval slice moves overall test EM by −0.014 and −0.011, and BM25 vanilla RAG from 0.138 to 0.113. The interval slice also posts the highest values on every retrieval-quality slice (TRP 0.376, CCR 0.920, and stale-blind CCR 0.773 where all other operators fall to 0.03–0.08), so per-operator retrieval comparisons involving interval are contaminated as well. **Evaluation-protocol fix.** Do not render `` (or otherwise expose `t_query`) for interval questions. Interval questions are self-contained: the reference event is stated in the question text, and the query year appears in the question text for only 2/864 interval items. When comparing against the reference-baseline numbers shipped with v1.0, either apply the same protocol or report EM with the interval slice excluded alongside overall EM. Do not report interval-slice accuracy obtained under a `t_query`-exposing protocol as a capability result. **What is not affected.** The gold supporting subgraphs (`S_star`), distractor subgraphs (`S_dist`), and stale subgraphs (`S_stale`) of interval questions are correct as released; the leak is a property of the evaluation-prompt protocol, not of the graph annotations. Question text, answers, and all non-interval operators are unaffected. ### 2. Stale negatives: construction property and intended use `S_stale` replaces one gold hop with the same `(s, r, o)` at a time invalid at `t_query`. Two properties of this construction should inform how the negatives are used: - **Excluded by a correct temporal filter, by definition.** The source KG is point-in-time (`t_start == t_end`), so validity at `t_query` reduces to exact-year equality, and a stale negative — required at construction to be invalid at `t_query` — is removed by any correctly implemented temporal filter. Empirically, 0 of the 1,429 functional stale negatives in the test split are valid at `t_query`, and the temporally-filtered reference retriever retrieves 0 of them (stale negative-hit-rate 0.000). This is a structural consequence, not a measured difficulty. Stale negatives should not be described as hard negatives for temporally-aware systems. - **100% BM25-tied: invisible to lexical ranking.** A stale negative and the gold hop it replaces differ only in the validity window, which is not a lexical feature. Under BM25 over `label(s) label(r) label(o)` documents, 1,429/1,429 (100.0%) of functional stale pairs score as exact ties. **Intended use.** The stale negatives are a diagnostic instrument for time-blind retrieval: a system that ignores temporal validity admits wrong-time evidence of the gold fact's type into 85.9% of retrievals (the exact released stale triple into 8.7%), while a temporally-filtered system admits none. They measure whether a retriever applies temporal filtering at all, not how well it ranks under temporal ambiguity. Per-question functional-negative flags ship with the release; stale-dependent evaluation should be restricted to the functional subset (1,429/1,743 test questions), and the interval × stale cell (n=12 functional) is too small to support slice-level claims. ### 2b. Distractor negatives: lexically hard, temporally easy `S_dist` swaps the object of one gold hop, keeping `(s, r)`. The two axes come apart, and both matter when reporting on this benchmark. - **Lexical ranking cannot separate them.** On the 1,237 functional distractor pairs in the test split, BM25 over `label(s) label(r) label(o)` scores gold above the released distractor in 489 cases (39.5%), *below* it in 472 (38.2%), and exactly tied in 276 (22.3%). Excluding ties, gold wins 489 of 961 — **50.9%, which is chance.** A purely lexical retriever has no signal here at all. - **A temporal filter removes most of them anyway.** 96.4% of released distractors are invalid at `t_query` and are dropped by the per-hop filter; only 3.6% survive it. This is a **side effect of construction** — `_build_distractor` relaxes the `valid_at` constraint when selecting the substitute object — not a designed temporal challenge. For the temporally-filtered reference retriever, typed distractor negative-hit-rate is 0.138. **Intended use.** Distractors are the residual difficulty for a time-aware system, and the honest framing is narrow: they show that answer-string identity is not evidence identity, and they defeat lexical retrieval outright. Do not describe them as temporally hard — the temporal filter disposes of 96.4% of them — and do not read the 3.6% that survive as a designed property. As with stale, restrict distractor-dependent evaluation to the functional subset using the shipped flags. --- ### 3. Reference retriever: full specification A reviewer noted that the BFS+BM25 reference retriever is underspecified on indexing and ranking. The paper has no room for it, so the complete spec is here. This describes the *reference baseline* whose outputs ship in `baselines/` — it is not part of the benchmark, and nothing in the benchmark depends on it. **Indexing.** One BM25 document per KG quadruple (550,376 documents). Document text is `label(s) label(r) label(o)`, lowercased and `\w+`-tokenised; unresolvable QIDs fall back to the raw QID string. **The validity window is never indexed** — time enters only through the graph filter, which is why a stale variant and its gold counterpart are lexically identical (§2). **BM25.** Okapi, `k1 = 1.5`, `b = 0.75`, with Lucene-style smoothed IDF `log(1 + (N - df + 0.5) / (df + 0.5))`, `N = 550,376`. Query terms are deduplicated (no query-term-frequency / `k3` component). `avgdl` is computed over the same corpus. **Candidate generation (BFS).** The anchor is the gold first-hop subject (oracle); the non-oracle variant uses a deliberately simple longest-substring linker over entity labels, 60.2% anchor recall. Expansion is undirected via the entity index, with a per-hop temporal filter `valid_at(t_query)` — exact-year equality, since the source KG is point-in-time. Per-hop caps are 60 / 200 / 400 for hops 1 / 2 / 3, applied in CSV load order **before** BM25 re-ranking. The pool is deduplicated on `(s, r, o)`, keeping the first time-version. **Ranking.** BM25 re-ranks the surviving pool. Candidates are grouped by `(s, r, o)`, rendered as one evidence line each, sorted by score descending, and truncated to `k = 15` (`k = 25` for the 3-hop ablation). The stale-blind ablation is the identical architecture with the temporal filter off. --- ### 4. Retrieval determinism: two rules a reimplementation must match Neither of these changes any published number, but both are load-bearing if you rebuild the reference retriever and expect to reproduce our figures. - **Tie-breaking is decided by rule, not by score.** As §2 records, 1,429/1,429 (100.0%) of functional stale pairs are *exact* BM25 ties, because the validity window is not a lexical feature. The tie-break rule therefore decides every stale-versus-gold ordering on its own. Ours breaks ties **reverse-lexicographically on the rendered evidence line**, everywhere. An earlier version of the metric scripts instead used a stable sort that preserved candidate-pool (CSV) load order; the two paths were reconciled to the deployed behaviour, and the reverse-lexicographic rule is what produced every number in the paper and in `baselines/`. - **The non-oracle anchor linker is deterministic.** The simple longest-substring linker over KG entity labels (gold-anchor recall 60.2%) was previously sensitive to `PYTHONHASHSEED` through set-iteration order. It is now deterministic. The published non-oracle 2-hop 3+-hop CCR moves **0.077 → 0.070** as a result. This is the same method made reproducible, **not** a correction to a result: no claim, comparison or conclusion depends on it. --- ## Regeneration **The benchmark itself is fully regenerable** from the code release (`code/`, stdlib-only, CPU-only). See `code/README.md`; you will need `tkgl-smallpedia_edgelist.csv` from TGB 2.0, which we do not redistribute. ``` python build_benchmark.py --kg_path --output_dir ./out/ \ --target_n 10000 --seed 42 ``` **The interval leak (§1) is checkable directly from the shipped data**, with no retriever and no extra scripts — for every interval question, `t_query` equals the gold answer year: ```python import json rows = [json.loads(l) for l in open("benchmark/benchmark_labelled.jsonl", encoding="utf-8")] iv = [r for r in rows if r["operator_type"] == "interval"] print(len(iv), sum(str(r["t_query"]).startswith(str(r["answer"])) for r in iv)) ``` **The retrieval-side numbers (§2–§4, and the paper's tables) are *not* regenerable from this release.** They require the reference retriever and generator, which are not included here. They are instead shipped as *outputs*: every figure behind those claims is in `baselines/`, so each is checkable without re-running anything. Split counts derive from the `split` field of `benchmark_labelled.jsonl`.