DUET

Dual Unlearning Evaluation across Training stages — a benchmark dataset for studying how fact salience and model training stage jointly affect machine unlearning in LLMs.

Introduced in Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning (arxiv:2602.19612) — accepted at ACL 2026 and ICLR 2026 Workshop.

The dataset contains Wikidata-derived question–answer pairs annotated with entity popularity scores (Wikipedia sitelinks + LLM confidence) and organized into paired forget/retain splits across two domains: general (mixed relations) and city (place-of-birth facts).

---

Quick start

from datasets import load_dataset

ds = load_dataset("SwetieePawsss/DUET")

forget = ds["forget_rare_10"]           # 10% forget, rare facts
retain = ds["retain_intersection_80"]   # paired retain set

city_forget = ds["city_forget_popular_5"]
city_retain = ds["city_retain_intersection_90"]

fast_retain = ds["city_fast_retain_500"]  # lightweight retain for fast eval

---

Fields

Field	Type	Description
question	string	Natural language question about a Wikidata triple
answer	string	Ground-truth answer
subjectLabel	string	Subject entity label
objectLabel	string	Object entity label
relation	string	Relation type (e.g. place_of_birth)
subject_qid	string	Wikidata QID of the subject
object_qid	string	Wikidata QID of the object
property_pid	string	Wikidata PID of the relation
pop_subject	float	Wikipedia sitelink count for the subject
pop_object	float	Wikipedia sitelink count for the object
pop_sum	float	Subject + object sitelinks — primary popularity signal
triple_pop	int	Discretized popularity bin index
best_gen_Llama_8b_Instract	string	Best Llama-3.1-8B-Instruct generation for this fact
gen_recall_Llama_8b_Instract	float	Token-level recall against the ground-truth answer
bert_sim_Llama_8b_Instract	float	BERTScore similarity to the ground-truth answer
source_csv	string	Source file identifier (corresponds to Wikidata relation)

Sample row

{
  "question": "What is the place of birth of Douglas Adams?",
  "answer": "Cambridge",
  "subjectLabel": "Douglas Adams",
  "relation": "place_of_birth",
  "objectLabel": "Cambridge",
  "subject_qid": "Q42",
  "property_pid": "P19",
  "object_qid": "Q350",
  "pop_subject": 174.0,
  "pop_object": 138.0,
  "pop_sum": 312.0,
  "triple_pop": 2,
  "best_gen_Llama_8b_Instract": "Cambridge",
  "gen_recall_Llama_8b_Instract": 1.0,
  "bert_sim_Llama_8b_Instract": 0.983,
  "source_csv": "place_of_birth"
}

---

Splits

Each forget set is paired with a retain set constructed as the intersection of the full corpus with the complement of the forget set.

General domain

Split	Examples	Description
full_	57,343	Full unfiltered corpus
filtered	28,634	Filtered subset used in benchmark experiments
forget_rare_1 / forget_popular_1	286	1% forget — rare / popular facts
retain_intersection_98	28,062	Retain set for 1% forget
forget_rare_5 / forget_popular_5	1,431	5% forget — rare / popular facts
retain_intersection_90	25,772	Retain set for 5% forget
forget_rare_10 / forget_popular_10	2,863	10% forget — rare / popular facts
retain_intersection_80	22,908	Retain set for 10% forget
forget_rare_500 / forget_popular_500	500	Fixed-size forget sets
retain_intersection_500	500	Fixed-size retain set

City domain (place-of-birth)

Split	Examples	Description
city_full	9,658	Full city subset
city_forget_rare_1 / city_forget_popular_1	96	1% forget — rare / popular city facts
city_retain_intersection_98	9,466	Retain set for 1% forget
city_forget_rare_5 / city_forget_popular_5	482	5% forget — rare / popular city facts
city_retain_intersection_90	8,694	Retain set for 5% forget
city_forget_rare_10 / city_forget_popular_10	965	10% forget — rare / popular city facts
city_retain_intersection_80	7,728	Retain set for 10% forget
city_fast_retain_500 / city_fast_retain_1500	500 / 1,500	Lightweight retain samples for fast evaluation
paraphrases_city_forget_rare_5	902	Paraphrased 5% forget — rare city facts
paraphrases_city_forget_popular_5	962	Paraphrased 5% forget — popular city facts

---

Citation

@misc{anna2026anatomyunlearningdualimpact,
  title={Anatomy of Unlearning: The Dual Impact of Fact Salience and Model Fine-Tuning},
  author={Borisiuk Anna and Andrey Savchenko and Alexander Panchenko and Elena Tutubalina},
  year={2026},
  eprint={2602.19612},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2602.19612}
}

---

License

MIT