Datasets:

yaopaul
/

DSEnt

license: cc-by-sa-4.0
task_categories:
  - question-answering
  - text-generation
language:
  - en
configs:
  - config_name: category1_forget
    data_files:
      - split: train
        path: category1_forget-train.json
  - config_name: category1_retain
    data_files:
      - split: train
        path: category1_retain-train.json
      - split: validation
        path: category1_retain-validation.json
      - split: test
        path: category1_retain-test.json
  - config_name: category2_forget
    data_files:
      - split: train
        path: category2_forget-train.json
  - config_name: category2_retain
    data_files:
      - split: train
        path: category2_retain-train.json
      - split: validation
        path: category2_retain-validation.json
      - split: test
        path: category2_retain-test.json
tags:
  - llm-unlearning
  - entity-level-unlearning
  - wikipedia
size_categories:
  - 1K<n<10K

DSEnt : Domain-Specific Entity Dataset

DSEnt is a domain-specific entity dataset built from English Wikipedia pages and automatically generated using GPT-5.5. The dataset spans multiple domains (categories). For each domain, it includes five real-world target entities, and each target entity is paired with five neighboring entities selected from related Wikipedia pages.

Category 1 : MLB Players.
Category 2 : Writers.

Usage in Python

from datasets import load_dataset

# Load forget data
train_data = load_dataset("yaopaul/DSEnt", name="category1_forget", split="train")

# Load retain data
retain_train_data = load_dataset("yaopaul/DSEnt", name="category1_retain", split="train")
retain_valid_data = load_dataset("yaopaul/DSEnt", name="category1_retain", split="validation")
retain_test_data = load_dataset("yaopaul/DSEnt", name="category1_retain", split="test")

Acknowledgement

The construction procedure of DSEnt was inspired by 6rightjade/ELUDe, particularly its LLM-assisted workflow for generating entity-level question-answer pairs. DSEnt extends this idea to domain-specific entity settings by collecting English Wikipedia-based target and neighboring entities across multiple different categories.