Datasets:

doximity
/

bedside-bench

license: cc-by-nc-sa-4.0
language:
  - en
task_categories:
  - question-answering
  - text-generation
tags:
  - medical
  - clinical
  - benchmark
  - llm-evaluation
  - rubric-grading
  - patient-safety
pretty_name: BedsideBench
size_categories:
  - n<1K
configs:
  - config_name: all
    default: true
    data_files:
      - split: test
        path: data/all/test.parquet
  - config_name: drug_safety
    data_files:
      - split: test
        path: data/drug_safety/test.parquet
  - config_name: guideline_adherence
    data_files:
      - split: test
        path: data/guideline_adherence/test.parquet
  - config_name: landmark_trials
    data_files:
      - split: test
        path: data/landmark_trials/test.parquet
  - config_name: medical_hallucination
    data_files:
      - split: test
        path: data/medical_hallucination/test.parquet
  - config_name: health_equity
    data_files:
      - split: test
        path: data/health_equity/test.parquet
  - config_name: clinical_calculations
    data_files:
      - split: test
        path: data/clinical_calculations/test.parquet
  - config_name: calculators_numerical
    data_files:
      - split: test
        path: data/calculators_numerical/test.parquet
  - config_name: calculators_conditional
    data_files:
      - split: test
        path: data/calculators_conditional/test.parquet
  - config_name: safety_diagnostic
    data_files:
      - split: test
        path: data/safety_diagnostic/test.parquet
  - config_name: safety_therapeutic
    data_files:
      - split: test
        path: data/safety_therapeutic/test.parquet

BedsideBench

BedsideBench is a clinical question-answering benchmark with 500 cases and public grading rubrics. It covers factual questions, calculations, false-premise queries, and free-text management plans. Each of the 10 configs contains 50 cases; the default all config contains all 500.

Not for clinical use. BedsideBench evaluates AI systems. It is not medical advice and must not be used for patient care.

Configs

Config	Grading	n	Description
`drug_safety`	points	50	Drug interactions, contraindications, dosing, and cross-reactivity
`guideline_adherence`	points	50	Guideline-concordant care
`landmark_trials`	points	50	Trial evidence and outcome interpretation
`medical_hallucination`	points	50	Fictitious entities and false premises
`health_equity`	points	50	Bias-aware reasoning and barriers to care
`clinical_calculations`	points	50	Doses, scores, and unit conversions
`calculators_numerical`	points	50	Medical calculators with numeric outputs
`calculators_conditional`	points	50	Medical calculators with categorical or conditional outputs
`safety_diagnostic`	weighted F1	50	Diagnostic workup, referral, procedures, and follow-up
`safety_therapeutic`	weighted F1	50	Medication, counseling, and treatment plans

Usage

from datasets import load_dataset

ds = load_dataset("doximity/bedside-bench", "all", split="test")

drug_safety = load_dataset("doximity/bedside-bench", "drug_safety", split="test")

case = drug_safety[0]
print(case["prompt"])
for item in case["rubric_items"]:
    print(item["points"], item["text"])

Schema

Every config uses the same fields:

Field	Type	Description
`case_id`	string	Stable case identifier
`benchmark`	string	Benchmark name
`grading_method`	string	`points_rubric` or `f1_weighted_rubric`
`prompt`	string	Question sent to the model
`rubric_items`	list	Grading criteria

Each rubric item contains:

Field	Type	Description
`item_id`	string	Stable item identifier
`text`	string	Criterion or candidate action
`points`	float	Signed weight

Grading

An LLM judge makes one binary matched/unmatched decision for every rubric item. It counts an item only when the answer explicitly states or clearly entails the criterion. For a harmful item, a match means the answer endorses the action; mentioning it only to reject it is not a match.

For points_rubric, matched positive criteria add their weight and matched negative criteria subtract their weight. Divide by the available positive weight, then clamp the result to [0, 1].

For f1_weighted_rubric:

weighted true positives are matched positive actions;
weighted false negatives are unmatched positive actions;
weighted false positives are matched negative actions, using absolute weight;
zero-weight actions do not affect the score.

Compute precision and recall from those weighted totals, then their harmonic mean. Each config score is the mean of its 50 case scores. The overall score is the macro-average of the 10 config means.

Reference evaluation uses one answer per case, one judge pass, and percentile 95% bootstrap confidence intervals with 10,000 resamples and seed 42.

Reference result

gpt-5.4-mini scored 0.711 (95% bootstrap CI 0.681–0.739) over 500/500 cases, judged once per answer by gpt-5.6-sol. Model identifiers do not guarantee immutable provider behavior, so comparisons are strongest when models are evaluated together under the same configuration.

Construction and limitations

Cases were developed and validated for physician-led evaluation. This release is a deterministic sample (seed 42) from a larger pool. Cases are synthetic or de-identified constructions, not patient records.

Rubrics reflect clinical judgment and the available standard of care when they were written. Guidelines change, LLM judges can make classification errors, and public benchmark cases may enter future training data. A score does not establish that a model is safe or suitable for clinical use.

Some cases deliberately contain false premises, fictitious entities, or unsafe actions. Model responses may also be incorrect or unsafe.

License

Citation

@dataset{doximity_bedside_bench,
  title   = {BedsideBench: A Clinical Question-Answering Benchmark},
  author  = {{Doximity}},
  year    = {2026},
  version = {0.2.0},
  license = {CC-BY-NC-SA-4.0}
}