AMNESIA / README.md
dsaeedeh's picture
Super-squash branch 'main' using huggingface_hub
747b4cb
|
Raw
History Blame
6.33 kB
metadata
license: cc-by-nc-sa-4.0
task_categories:
  - question-answering
language:
  - en
tags:
  - machine-unlearning
  - medical
  - clinical
  - privacy
  - benchmark
pretty_name: AMNESIA
size_categories:
  - 10K<n<100K
configs:
  - config_name: random
    data_files:
      - split: forget_5
        path: forget_5_with_category.jsonl
      - split: forget_10
        path: forget_10_with_category.jsonl
      - split: forget_15
        path: forget_15_with_category.jsonl
      - split: forget_20
        path: forget_20_with_category.jsonl
      - split: forget_25
        path: forget_25_with_category.jsonl
      - split: retain_5
        path: retain_5_with_category.jsonl
      - split: retain_10
        path: retain_10_with_category.jsonl
      - split: retain_15
        path: retain_15_with_category.jsonl
      - split: retain_20
        path: retain_20_with_category.jsonl
      - split: retain_25
        path: retain_25_with_category.jsonl
      - split: hold_out
        path: hold_out_with_category.jsonl
      - split: rest
        path: rest_with_category.jsonl
  - config_name: cancer
    data_files:
      - split: forget_5
        path: cancer_forget_5_with_category.jsonl
      - split: forget_10
        path: cancer_forget_10_with_category.jsonl
      - split: forget_15
        path: cancer_forget_15_with_category.jsonl
      - split: forget_20
        path: cancer_forget_20_with_category.jsonl
      - split: forget_25
        path: cancer_forget_25_with_category.jsonl
      - split: retain_5
        path: cancer_retain_5_with_category.jsonl
      - split: retain_10
        path: cancer_retain_10_with_category.jsonl
      - split: retain_15
        path: cancer_retain_15_with_category.jsonl
      - split: retain_20
        path: cancer_retain_20_with_category.jsonl
      - split: retain_25
        path: cancer_retain_25_with_category.jsonl
  - config_name: cardiovascular
    data_files:
      - split: forget_5
        path: cardiovascular_forget_5_with_category.jsonl
      - split: forget_10
        path: cardiovascular_forget_10_with_category.jsonl
      - split: forget_15
        path: cardiovascular_forget_15_with_category.jsonl
      - split: retain_5
        path: cardiovascular_retain_5_with_category.jsonl
      - split: retain_10
        path: cardiovascular_retain_10_with_category.jsonl
      - split: retain_15
        path: cardiovascular_retain_15_with_category.jsonl
  - config_name: infectious_or_immune_diseases
    data_files:
      - split: forget_5
        path: infectious_or_immune_diseases_forget_5_with_category.jsonl
      - split: forget_10
        path: infectious_or_immune_diseases_forget_10_with_category.jsonl
      - split: forget_15
        path: infectious_or_immune_diseases_forget_15_with_category.jsonl
      - split: forget_20
        path: infectious_or_immune_diseases_forget_20_with_category.jsonl
      - split: retain_5
        path: infectious_or_immune_diseases_retain_5_with_category.jsonl
      - split: retain_10
        path: infectious_or_immune_diseases_retain_10_with_category.jsonl
      - split: retain_15
        path: infectious_or_immune_diseases_retain_15_with_category.jsonl
      - split: retain_20
        path: infectious_or_immune_diseases_retain_20_with_category.jsonl

AMNESIA: A Large-Scale Medical Unlearning Benchmark Suite with Disease-Informed Analysis

AMNESIA is the first large-scale, open-source benchmark for machine unlearning in the medical domain. It is designed to evaluate methods that selectively remove the influence of specific patient data from trained medical language models while preserving the model's overall clinical knowledge and utility — a capability motivated by privacy regulations such as the GDPR "right to be forgotten."

The benchmark is built on de-identified clinical notes drawn from PubMed Central and paired with model-answerable question–answer items, organized into forget/retain partitions at multiple removal ratios and across disease categories.

Dataset structure

Each line of every .jsonl file is a single JSON record built from a de-identified patient note. A record contains:

  • a de-identified clinical patient note from PubMed Central,
  • factual questions testing direct recall from the note,
  • reasoning questions requiring multi-hop clinical inference,
  • gold-standard answers as verified references, and
  • a disease category label (one of the CDC-classified disease types).

Splits

The suite provides two kinds of partitioning, exposed here as loadable configurations:

Random (patient-level) partitioning. Patients are selected at random to form forget sets at 5%, 10%, 15%, 20%, and 25%; the remaining patients form the corresponding retain sets. hold_out contains held-out QA items that are never seen during training or unlearning, used to test generalization. rest holds the remaining data.

Disease-level partitioning. The same forget/retain ratios are applied within individual disease cohorts — Cancer, Cardiovascular, and Infectious/Immune Diseases — enabling disease-informed analysis. (Note that the Cardiovascular cohort provides the 5/10/15% ratios, while Cancer and Infectious/Immune provide additional higher ratios, reflecting cohort sizes.)

The forget sets are nested: each forget_X is a subset of the next larger forget_(X+5), which makes evaluations across removal ratios directly comparable and reproducible.

Usage

from datasets import load_dataset

# Random patient-level splits
ds = load_dataset("dsaeedeh/AMNESIA", "random", split="forget_10")

# Disease-level splits
cancer = load_dataset("dsaeedeh/AMNESIA", "cancer", split="forget_10")
cvd    = load_dataset("dsaeedeh/AMNESIA", "cardiovascular", split="retain_5")

You can also stream individual files directly:

ds = load_dataset("json", data_files="hold_out_with_category.jsonl")

Citation

If you use AMNESIA, please cite:

@article{davoudi2026amnesia,
  title   = {AMNESIA: A Large-Scale Medical Unlearning Benchmark Suite with Disease-Informed Analysis},
  author  = {Davoudi, Saeedeh and Iranmanesh, Reihaneh and Frieder, Ophir and Goharian, Nazli},
  journal = {arXiv preprint arXiv:2605.30599},
  year    = {2026}
}

License

Released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0) license. Non-commercial use only, with attribution and share-alike.