README.md

metadata

license: cc-by-4.0
configs:
  - config_name: raw
    data_files:
      - split: raw_forget
        path: raw/raw_forget.parquet
      - split: raw_holdout
        path: raw/raw_holdout.parquet
      - split: raw_retain1
        path: raw/raw_retain1.parquet
      - split: raw_retain2
        path: raw/raw_retain2.parquet
  - config_name: knowmem
    data_files:
      - split: forget_qa
        path: knowmem/forget_qa.parquet
      - split: retain_qa
        path: knowmem/retain_qa.parquet

Details on Subsets & Splits

• raw: A raw corpus from which all subsets are derived. The splits are:
  • raw_forget: Data intended to be forgotten
  • raw_holdout: Data never seen by the target model during pre-training and unlearning
  • raw_retain1: Data used optionally as a calibrator for unlearning
  • raw_retain2: Retain set, i.e. data seen by the target model and used for evaluation