saxenan3/temporalwiki-drift-cl-easy

TemporalWiki Drift — Continual-Learning Benchmark (LLM-Curated easy Variant)

A higher-signal variant of the TemporalWiki Drift CL benchmark. The original used <subject> <relation> string-concat prompts on raw snapshot-diff triples; this variant filters editor noise via an LLM judge and re-expresses every triple as a closed-book natural-language Q/A.

Splits

split	rows	description
train	~1500	curated drift triples to learn, ~500 per slice; some rows are paraphrase Q's on the same fact to fill the row target
test	150	held-out paraphrase Q's, 50 per slice (same triples as train, paraphrase wording — phrasing-robust acquisition probe)
stable	50	stable facts about the same train subjects (corruption probe)

Processing pipeline

Source: monthly Wikipedia + Wikidata snapshots at 2025-11-20, 2025-12-01, 2026-01-01, 2026-02-01. Per adjacent pair we extract changed.csv (subject, relation, object) where the object value differs across the pair.

1. Layer 1 — Rule allowlist (≈38 relations). Keep relations that encode real time-varying events: career roles (head coach, employer, position held, head of state/government, mayor/governor/CEO/ chairperson/secretary general/ambassador/chancellor/manager), life events (spouse, country of citizenship, convicted of, military rank, nominated for, award received), corporate M&A (owned by, owner of, parent organization, subsidiary, headquarters location, founded by), media production (cast member, voice actor, screenwriter, composer, production company), child, significant event. Drop relations with high editor-noise rate (participant, participant in, residence).

2. Layer 2 — LLM relation calibration. Sample 10 random triples per relation, judge each KEEP/DROP with Qwen3-8B. Drop the whole relation if KEEP_rate < 0.55.

3. Layer 3 — LLM triple judge (3-of-3 strict). For each (subject, relation) in the 3-slice intersection (objects distinct in every slice), ask Qwen3-8B KEEP/DROP for every slice's version of the triple. Keep only keys where all 3 judgments are KEEP. The judge prompt uses few-shot examples that explicitly reject:

   • subject = historical event (e.g. 2012 UEFA CL Final — participants are permanent, not 2025 drift);
   • stable lifelong attributes (field of work of historical figures);
   • editor cycling among equally-valid values (e.g. career-long awards);
   • synonym/hypernym swaps; vandalism reverts.

   The judge defaults to DROP when it cannot verify the change occurred in the snapshot window (false DROPs are cheaper than false KEEPs).

4. Natural-language Q/A generation. For each surviving triple, fetch the Wikipedia article body at snapshot_new from the per-snapshot SQLite mirror and prompt Qwen3-8B (/no_think) to write a closed-book question whose unique answer is exactly the gold value. The Q must name the subject fully and contain an explicit as of <snapshot_new> date anchor. A separate self-consistency call answers the Q from the article alone; the resulting F1 (qa_self_consistency_f1) is stored per row for post-hoc confidence-filtering — it is not used to drop rows (multi-valued relations would over-reject otherwise).

5. Auto-expansion via multi-paraphrase. If the L3 survivor pool is smaller than the target n_train=500 per slice, extra paraphrase Q's are generated on the first surviving keys to fill the gap (IMPRINT-style multi-paraphrase-per-fact training). The unique-fact count is recorded in the run manifest.

6. Same-subject stable. Stable probes are drawn from unchanged.csv rows about the same subjects that appear in train (corruption probe on entities the model just trained on). If insufficient same-subject stable rows pass Q/A generation, the script falls back to the original variant's stable subjects.

Columns

Every row carries both representations + the LLM-generated NL Q:

• subject, relation, object — the factual triple (training target)
• question — NEW: the natural-language question (the user prompt)
• article — full Wikipedia article body at snapshot_new (for retrieval / KV-cache methods)
• subject_sitelink — canonical Wikipedia title used for article lookup
• slice, slice_tag, snapshot_new, snapshot_old, data_source — provenance
• qa_style — "direct" / "paraphrase" / "train_paraphrase" / "fallback_direct"
• qa_self_consistency_f1 — RAG-oracle F1 (the Qwen3-8B answer to the question, given the article, vs gold). Higher = more confident the Q+gold are well-formed
• qa_article_chars — article length the Q was conditioned on
• judge_passes — 3 = passed strict 3-of-3 KEEP; 0 = stable (no judge needed)
• continuous_reward — flag opting into continuous-F1 reward (vs binary hit threshold)

Snapshots

2025-11-20, 2025-12-01, 2026-01-01, 2026-02-01. Drift slices are the three adjacent pairs; stable facts are sampled to be invariant across all four.

Relation to original variant

Both variants share the (subject, relation, object) backbone and overlapping provenance columns, so existing trainers and eval pipelines pick up the easy variant by swapping a data path. The easy variant adds the question, qa_*, judge_passes, and continuous_reward columns to support NL-Q/A training and post-hoc quality filtering.