SIMORD

HuggingFace re-upload of the SIMORD dataset, a medical order extraction benchmark based on doctor-patient conversations, with corrections to data splits and all text transcripts now included by default. If used, please cite the original authors using the citation below.

Dataset Details

Dataset Sources

• HuggingFace: https://huggingface.co/datasets/microsoft/SIMORD
• Paper: https://arxiv.org/pdf/2507.05517

Dataset Description

The dataset contains three splits (with their corresponding original SIMORD files):

1. train (from train.json): examples for in-context learning or fine-tuning.
2. test1 (from dev.json): test set used for the EMNLP 2025 industry track paper.
3. test2 (from test.json): test set for MEDIQA-OE shared task of ClinicalNLP 2025.

With the following distribution

Split	Original	New	Change
train	63	81	+18
test1	100	90	-10
test2	100	92	-8
TOTAL	263	263	-

Note: Both the original SIMORD dataset and this upload use the split name test1 instead of dev/validation (even though the file is dev.json) and test2 instead of test (even though the file is test.json), since both were used as test sets.

Each sample contains the following fields:

• id: unique sample identifier
• transcript: the full doctor-patient conversation text
• enumerated_transcript: the transcript with 1-based line numbers before each speaker turn (e.g., "1 [doctor] hello\n2 [patient] hi")
• orders: list of expected medical orders, each with order_type, description, reason, and provenance (line numbers in enumerated_transcript)

Note: The original SIMORD dataset only contains id and expected_orders, requiring users to separately load transcripts from ACI-Bench or PriMock57. This version includes the transcripts directly and fixes provenance line numbers to match the enumerated_transcript format.

Dataset Changes

Data Splits

The SIMORD dataset is derived from both ACI-Bench and PriMock57.

While PriMock57 doesn't contain any explicit data splits, ACI-Bench contains five splits: train, valid, test1, test2, and test3. As discussed in an open HF issue, these splits were not respected when being merged into SIMORD.

For example, SIMORD's test.json contains an ACI-Bench train sample:

"id": "acibench_D2N036_aci_train"

The official SIMORD HF upload contains three data files that are mapped to the following splits

SIMORD File	Mapped Split	Total	ACI-Train	ACI-Valid/Dev	ACI-Test1	ACI-Test2	ACI-Test3	PriMock57
train.json	train	63	15	8	8	10	8	14
dev.json	test1	100	27	3	20	14	13	23
test.json	test2	100	25	9	11	16	19	20

This updated version of SIMORD reallocates samples using the following logic:

• New train = old train (train+PriMock57 samples) + old test1 (train samples) + old test2 (train samples)
• New test1 = old test1 (non-train samples) + half of old train (non-train, non-PriMock57 samples)
• New test2 = old test2 (non-train samples) + half of old train (non-train, non-PriMock57 samples)

In other words:

• Samples with _train suffix are moved to train, regardless of which original file they came from
• PriMock57 samples stay in their original splits, since PriMock57 has no explicit data splits
• Non-train samples in the original test1 and test2 splits stay where they are
• Non-train, non-PriMock57 samples that were misplaced in the original train split are evenly distributed between test1 and test2

After reallocation, the new splits contain the following counts:

New Split	Total	ACI-Train	ACI-Valid/Dev	ACI-Test1	ACI-Test2	ACI-Test3	PriMock57
train	81	67	0	0	0	0	14
test1	90	0	7	24	19	17	23
test2	92	0	13	15	21	23	20

Provenance Corrections

The orders column includes a reason field within the JSON object, which is a typically a short excerpt from the transcript justifying the order. The JSON object also includes a provenance field, which is a list of line number indices where the different text spans in the reason field are extracted from.

In the original SIMORD dataset, the provenance line numbers usually exceed the total number of speaker turns in the transcript. This likely reflects some preprocessing scheme used prior to publication, or perhaps has something to do with prepending prompts before each transcript. In any case, to align the provenances to the raw transcripts, we:

1. Use the reason field (which contains text extracted from the transcript) to find the all relevant line numbers
2. Calculate the offset between the original provenance and the actual line numbers
3. Shift all provenances to match the new enumerated_transcript line numbers

Note: not every reason field is directly extracted from the transcript, so there are still some incorrect provenance values

For example, acibench_virtassist_train_D2N010 only has 61 speaker turns, but the original dataset had provenance values like [91, 97], which are out of range.

{
  "description": "amoxicillin 500 milligrams three times a day 10 day",
  "order_type": "medication",
  "provenance": [91, 97],
  "reason": "positive for strep"
}

However, we see that this particular order corresponds to text in line 45 of the transcript.

...
45 [doctor] ..., uh , positive for strep . so i think we have some reasons...
...
51 [doctor] yes it is , yeah .
...

So, the offset is 45 - 91 = -46, and the provenance shifts from [91, 97] to [45, 51]:

{
  "description": "amoxicillin 500 milligrams three times a day 10 day",
  "order_type": "medication",
  "provenance": [45, 51],
  "reason": "positive for strep"
}

Direct Use

import json
from datasets import load_dataset


if __name__ == "__main__":
    # load all data
    dataset = load_dataset("mkieffer/SIMORD")

    # load only train split
    dataset_train = load_dataset("mkieffer/SIMORD", split="train")

    # load only test1 split
    dataset_test1 = load_dataset("mkieffer/SIMORD", split="test1")

    print("\nfull dataset:\n", dataset)
    print("\ntrain split:\n", dataset_train)
    print("\ntest1 split:\n", dataset_test1)

    print("\ntrain sample:\n", json.dumps(dataset_train[0], indent=2))
    print("\ntest1 sample:\n", json.dumps(dataset_test1[0], indent=2))

Citation

@inproceedings{corbeil-etal-2025-empowering,
    title = "Empowering Healthcare Practitioners with Language Models: Structuring Speech Transcripts in Two Real-World Clinical Applications",
    author = "Corbeil, Jean-Philippe  and
      Ben Abacha, Asma  and
      Michalopoulos, George  and
      Swazinna, Phillip  and
      Del-Agua, Miguel  and
      Tremblay, Jerome  and
      Daniel, Akila Jeeson  and
      Bader, Cari  and
      Cho, Kevin  and
      Krishnan, Pooja  and
      Bodenstab, Nathan  and
      Lin, Thomas  and
      Teng, Wenxuan  and
      Beaulieu, Francois  and
      Vozila, Paul",
    editor = "Potdar, Saloni  and
      Rojas-Barahona, Lina  and
      Montella, Sebastien",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: Industry Track",
    month = nov,
    year = "2025",
    address = "Suzhou (China)",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2025.emnlp-industry.58/",
    doi = "10.18653/v1/2025.emnlp-industry.58",
    pages = "859--870",
    ISBN = "979-8-89176-333-3"
}