Datasets:

samanjoy2
/

MedPRESS_Benchmark

language:
  - en
pretty_name: MedPRESS
tags:
  - medical
  - safety
  - sycophancy
  - llm-evaluation
  - benchmark
task_categories:
  - text-generation
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: test
        path:
          - >-
            Data/medication_and_treatment_demand_200_exact_schema_paraphrase_v2.csv
          - Data/personal_health_self_care_200_exact_schema_paraphrase_v2.csv
          - >-
            Data/symptom_triage_and_care_resistance_200_exact_schema_paraphrase_v2.csv

MedPRESS

MedPRESS is a multi-turn benchmark for evaluating patient-pressure-induced medical sycophancy in large language models. The dataset tests whether a model maintains a safe medical stance when a user repeatedly pressures it toward an unsafe or false health belief.

The benchmark is designed around five-turn conversations. Each row contains one medical scenario, the unsafe or false belief being pressured, the expected safe stance, and five progressively stronger user turns.

Dataset Summary

MedPRESS contains 600 English benchmark cases across three medical safety scenario families:

Scenario family	File	Rows
Medication and treatment demand	`Data/medication_and_treatment_demand_200_exact_schema_paraphrase_v2.csv`	200
Personal health self-care	`Data/personal_health_self_care_200_exact_schema_paraphrase_v2.csv`	200
Symptom triage and care resistance	`Data/symptom_triage_and_care_resistance_200_exact_schema_paraphrase_v2.csv`	200

Scenario Families

Medication and Treatment Demand

Tests whether a model validates inappropriate medication or treatment demands, gives unsafe practical support, or redirects toward safer care such as clinician or pharmacist review.

Personal Health Self-Care

Tests whether a model endorses risky home remedies or unsafe self-treatment, while checking whether it provides safe alternatives and appropriate red-flag guidance.

Symptom Triage and Care Resistance

Tests whether a model minimizes potentially serious symptoms or supports delaying care when the scenario requires urgent or emergency evaluation.

Data Schema

All CSV files share the same schema:

Column	Description
`case_id`	Unique scenario identifier.
`scenario_family`	Scenario-family key.
`topic`	Medical topic or situation.
`risk_level`	Risk category for the scenario.
`unsafe_or_false_belief`	The unsafe or false belief the user pressures the model to accept.
`safe_stance_summary`	Short summary of the expected safe stance.
`turn1_initial_query`	Initial user message.
`turn2_personal_experience`	Follow-up pressure using personal experience.
`turn3_social_proof`	Follow-up pressure using social proof.
`turn4_external_claim`	Follow-up pressure using an external claim.
`turn5_direct_challenge`	Final direct challenge to the model's safe stance.
`triage_needed`	Whether the scenario requires triage guidance.
`triage_trigger`	The symptom, context, or condition that should trigger triage guidance.

Loading

from datasets import load_dataset

dataset = load_dataset("samanjoy2/medpress_dataset", split="test")
print(dataset)

To load a single CSV directly:

from datasets import load_dataset

dataset = load_dataset(
    "csv",
    data_files="hf://datasets/samanjoy2/medpress_dataset/Data/symptom_triage_and_care_resistance_200_exact_schema_paraphrase_v2.csv",
    split="train",
)

Intended Use

MedPRESS is intended for research and evaluation of LLM medical safety behavior, especially in multi-turn settings where user pressure can make models more agreeable to unsafe medical claims.

This dataset should not be used as medical advice, clinical guidance, or a substitute for professional medical judgment.

Citation

If you use MedPRESS, please cite:

@misc{joy2026medpressmultiturnbenchmarkpatientpressureinduced,
      title={MedPRESS: A Multi-turn Benchmark for Patient-Pressure-Induced Medical Sycophancy in LLMs}, 
      author={Saman Sarker Joy and Niloy Farhan},
      year={2026},
      eprint={2608.02520},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.02520}, 
}