README.md

---
language:
- en
license: cc-by-4.0
task_categories:
- text-generation
pretty_name: GRAND-ROUNDS
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---
# GRAND-ROUNDS

GRAND-ROUNDS (Graded Responses and Annotated Notes for Diagnostic Reasoning on UNstructured Data Sets)
is a physician-annotated benchmark for validating LLM judges of open-ended clinical reasoning.
GRAND-ROUNDS comprises 9,217 physician scores across 5,250 scored response entries from 160
clinicians and 9 AI models across six tasks, drawn from seven published studies and graded by
11 physicians. This release contains all of GRAND-ROUNDS except the BIDMC emergency-department
task (Brodeur et al. 2026), which contains protected health information: 4,339 entries with
7,395 independent physician scores across the five remaining tasks. Released with the paper
*Scaling Clinical Judgment to Evaluate Medical AI*. Website: https://preceptron.net · Code: https://github.com/2v/PrecepTron ·
Models: https://huggingface.co/collections/tbuckley/preceptron-6a3194798cea4c5d6bc9713f

## Benchmarks

| Benchmark | Description | Records |
|-----------|-------------|---------|
| `management_reasoning` | Grey Matters management cases (Goh et al. 2025): free-text management answers scored on case-specific rubrics | 2,765 |
| `cpc_bond` | NEJM clinicopathological conference (CPC) cases: differential diagnoses scored with the 0-5 Bond score | 853 |
| `r_idea` | NEJM Healer cases (Cabral et al. 2024): clinical reasoning documentation scored with the 10-point R-IDEA rubric | 312 |
| `diagnostic_reasoning` | Landmark diagnostic cases (Goh et al. 2024): structured diagnostic reasoning scored on the 19-point rubric | 278 |
| `cpc_management` | NEJM CPC cases: proposed diagnostic testing plans scored on the 0-2 testing-plan rubric | 131 |

## Files

- `data/train-*.parquet`: the benchmark, loaded by `datasets.load_dataset("tbuckley/GRAND-ROUNDS")`.
- `combined_dataset.json`: the same records as one JSON list. The code at
  https://github.com/2v/PrecepTron reads this file from `score_data/combined_dataset.json`;
  the public repo does not ship it, so fetch it there with
  `hf download tbuckley/GRAND-ROUNDS combined_dataset.json --repo-type dataset --local-dir score_data`
  after `hf auth login`.

## Common Fields

- `benchmark`: Source benchmark identifier
- `case_id`: Case identifier
- `model`: AI model or human participant group
- `study`: Source study citation key
- `response`: Free-text response from model or participant
- `grade`: JSON string — list of grader scores (`grader`, `score`, and benchmark-specific fields)
- `final_diagnosis`: Ground truth diagnosis (where applicable)

Fields present only in some benchmarks: `question_number`, `question_text`, `max_score`,
`run_number`, `participant_id`, `asked_together`, `aliquot`, `cannot_miss_diagnoses`,
`cannot_miss_score`, `dataset`, `questions_raw`, `test_plan`, `case_vignette_multi`.

`case_vignette_multi` (management_reasoning only) is a JSON string containing an
ordered list of aliquots, each `{"vignette": "<cumulative case text visible at this
point>", "questions": ["q1", ...]}`. Use it to recover the incremental-disclosure
vignette that was visible when each question was asked; the flat `case_vignette`
field remains the full concatenation of every aliquot for backward compatibility.

## Case text availability

Full case text (`case_vignette`) is included for `management_reasoning` (the Grey Matters
cases), released here for the first time. The Landmark Diagnostic Cases (`diagnostic_reasoning`)
ship with responses, physician scores, and final diagnoses but not the case presentations, and
the NEJM CPC and NEJM Healer case presentations are copyrighted by the publisher, so
`diagnostic_reasoning`, `cpc_bond`, `cpc_management`, and `r_idea` carry no case text
(`is_case_released` is `False` on those rows). Note that the PrecepTron judge for
`diagnostic_reasoning` reads the case vignette, so that task's judge benchmark cannot be
re-run from this release alone. The BIDMC emergency-department task reported in the paper
is not included (patient-derived data).

## Citation

```bibtex
@article{buckley2026preceptron,
  title   = {Scaling Clinical Judgment to Evaluate Medical AI},
  author  = {Buckley, Thomas A. and Kanjee, Zahir and Brodeur, Peter G. and
             Crowe, Byron and Pettinato, Anthony M. and Shah, Aashna P. and
             Haimovich, Adrian D. and McCoy, Liam G. and Restrepo, Daniel and
             Freed, Jason A. and Goh, Ethan and Chen, Jonathan H. and Zwaan, Laura and
             Goodman, Katherine E. and Morgan, Daniel J. and
             Abdulnour, Raja-Elie E. and Rodman, Adam and Manrai, Arjun K.},
  year    = {2026},
  journal = {arXiv preprint arXiv:2609.12822},
  url     = {https://arxiv.org/abs/2609.12822}
}
```