GENIE: GENeralisation In Edits

A benchmark for model editing that asks whether an edited language model can use a new fact, not just recall it.

Each case is one counterfactual edit (subject, relation, old object → new object) built from Wikidata, paired with up to ten probes that move progressively further from the edit prompt: reworded, chained through related facts, and finally abstracted so the subject is no longer named. If an edit has changed what the model knows, the new fact should surface at every rung.

Built for the MSc thesis Model editing: capacities and limits (Ben Bethell, Birkbeck, University of London, 2026).

• Code and construction pipeline: github.com/benji-benji/genie-benchmark
• Thesis code: github.com/benji-benji/BBK-model-editing-final-report

Usage

from datasets import load_dataset

genie = load_dataset("benji-benji/genie", "full", split="test")          # 7,340 cases (v2)
gpt2 = load_dataset("benji-benji/genie", "gpt2-xl", split="test")        # GPT-2-XL set (v2)
thesis = load_dataset("benji-benji/genie", "v1-gpt2-xl", split="test")   # as evaluated in the thesis

Configurations

v2 is the recommended version. v1 is the data exactly as evaluated in the thesis.

Config	Cases	Probes	Contents
full	7,340	48,533	Full benchmark, 43 relations (v2)
gpt2-xl	67	465	GPT-2-XL evaluation set (v2)
llama3-8b	375	2,771	LLaMA-3-8B evaluation pool (v2)
v1-full	7,340	48,795	Full benchmark, as evaluated
v1-gpt2-xl	95	680	The 95 GPT-2-XL cases evaluated in the thesis
v1-llama3-8b	390	2,880	The pool from which the thesis sampled 95 LLaMA-3-8B cases

Data fields

Field	Description
case_id	Unique id, <relation>_<n>
subject	The entity being edited
relation	Wikidata relation name
prompt	Edit prompt with {} for the subject
target_old / target_new	Original object / counterfactual object
meta	Wikidata property id (pid), subject frame (work, person, org, place), year, location, sitelinks
probes	List of probes: category, prompt, answer_old, answer_new (accepted answers), optional constraints

Probe categories

Probe	What it asks
efficacy	The edit prompt itself
paraphrase_1–3	The same question, reworded at increasing length and indirection
hop_1–3	A chained question about the edited object, one to three properties deep (e.g. the birthplace of the composer of E.T.)
abstract_1	A forced choice between two other entities: one that shares the old object, one that shares the new
abstract_2	Open generation one rung up: the subject is named, but the relation is described in more general terms
abstract_3	Open generation with the subject removed, identified only by constraints such as year and a related person

Probe composition is ragged by design: a probe is only generated when Wikidata supports it. abstract_2 and abstract_3 have no gold continuation and are scored by generating 20 tokens and checking whether an accepted answer appears as a whole word.

Construction

1. Seed subjects: the most-viewed Wikipedia pages (Wikimedia pageviews API) and notable works (Wikidata SPARQL).
2. Facts, a counterfactual object of similar era and prominence, comparators and hop chains: Wikidata.
3. Probes generated from templates per relation and subject frame.
4. Per-model sets: each model was run once, unedited, over every abstract_2 and abstract_3 probe. Probes that never produced the old answer are dropped, then only cases that still hold an abstract_3 probe are kept.

Limitations

• Counterfactual objects are chosen to be plausible (same type, similar era and prominence) but are not checked by hand.
• Templates are English only and were written for the relations and frames sampled; some relations have no abstract_2 or abstract_3 templates.
• The per-model sets depend on the surfacing run for that model and are not transferable to other models; rerun the filter for a new model.
• Accepted answers are Wikidata labels plus simple variants (e.g. surname only); correct answers phrased differently are scored as misses.

Licence

CC BY 4.0. Underlying facts from Wikidata (CC0 1.0). Relation identifiers and popularity sampling adapted from RippleEdits (MIT).

Citation

Bethell, B. (2026) Model editing: capacities and limits. MSc Data Science thesis,
Birkbeck, University of London.