README.md

---
language:
- en
license: cc-by-4.0
tags:
- medical
- robustness
- benchmark
- arxiv:2610.08585
pretty_name: MedDistractNotes
configs:
- config_name: pairs
  data_files:
  - split: test
    path: data/ambient/pairs_v3.parquet
  default: true
- config_name: insertions
  data_files:
  - split: test
    path: data/ambient/insertions_v3.parquet
- config_name: organ_systems
  data_files:
  - split: test
    path: data/ambient/encounter_systems.parquet
- config_name: notes
  data_files:
  - split: test
    path: results/raw_v3/amb_notes.parquet
- config_name: judgments
  data_files:
  - split: test
    path: results/raw_v3/amb_judgments.parquet
- config_name: attribution
  data_files:
  - split: test
    path: results/raw_v3/amb_attribution.parquet
- config_name: single_note_judgments
  data_files:
  - split: test
    path: results/judge_control/judgments.parquet
- config_name: failure_modes
  data_files:
  - split: test
    path: results/v3/amb_failure_notes.parquet
---

# MedDistractNotes

MedDistractNotes tests whether incidental conversation enters clinical notes. It contains **1,152 clean–distracted transcript pairs from 576 encounters**, with one bystander and one nonliteral aside per encounter. The clean and distracted transcripts differ only by the inserted exchange. Use the paired inputs to evaluate your own note-generation model; the additional configurations contain the frozen study outputs.

```python
from datasets import load_dataset
pairs = load_dataset("NYU-OLAB/MedDistractNotes", "pairs", split="test")
row = pairs[0]
print(row["clean_transcript"], row["distracted_transcript"])
```

The `test` split is the entire evaluation benchmark. `source_dataset` preserves the original corpus split; these are not new training/development partitions. The same encounter appears in both perturbation families, so split by encounter if making your own partitions.

## Contents and keys

| Configuration | Rows | Contents |
|---|---:|---|
| pairs | 1,152 | Clean/distracted transcripts, reference note, inserted exchange, target summary and insertion position |
| insertions | 1,152 | Frozen GPT-generated exchanges, assigned content, organ system and generation metadata |
| organ_systems | 577 | Labels/rationales for the original cohort, including the one excluded single-speaker encounter |
| notes | 18,432 | Saved clean/distracted notes from eight models |
| judgments | 18,432 | Paired contamination/severity judgments and note-quality scores |
| attribution | 3,921 | Attribution and clinical-use classifications for flagged notes |
| single_note_judgments | 18,432 | Matched single-note control scores and explanations |
| failure_modes | 9,216 | Frozen outcome, failure-mode and section-location annotations for distracted notes |

Pair keys are `source_dataset`, `item_id`, `distractor_type`. Add `model` and `condition` to join per-note results. `model` retains the identifiers used in the frozen runs. Judge explanations are model-generated, not clinical adjudications. Two empty saved note records remain in the primary denominator; absent and unresolved attribution classifications are retained. Missing scores must not be treated as zero.

`protocol/prompts.json` contains the generation and scoring templates. Source data, outputs and scores are unchanged from the checked submission archive; only provider bookkeeping was removed from the single-note control export. `release_manifest.json` records content hashes and `data_schema.json` describes every column. The toolkit supports downloading these files into the expected analysis paths.

## Sources and license

Derived from [ACI-Bench](https://github.com/wyim/aci-bench) (Yim et al., 2023) and [MTS-Dialog](https://github.com/abachaa/MTS-Dialog) (Ben Abacha et al., 2023), both under CC BY 4.0. The released cohort uses 77 ACI-Bench encounters and 499 MTS-Dialog excerpts. The original transcripts and reference notes are retained; the changes are generated incidental exchanges and the associated model outputs/annotations. Cite the source corpora as well as this benchmark. This release uses CC BY 4.0 with source attribution retained.

## Intended use and limitations

For research on contamination, attribution and robustness in note generation. The asides are synthetic and the reference notes come from the source corpora. The scores are automated judgments, not clinician-validated safety labels. High general note-quality scores do not establish that a note is free of contamination. Do not use the benchmark as patient-care guidance.

Code: https://github.com/nyuolab/llm_distract (the public toolkit release is being prepared).

## Citation

[Paper](https://arxiv.org/abs/2610.08585). The citation below uses the publicly posted v1 title; the authors have submitted a revised title. The scientific data are frozen to the submitted study.

```bibtex
@misc{vishwanath2026incidental,
  title={Incidental information contaminates patient notes and disrupts clinical reasoning in large language models},
  author={Vishwanath, Krithik and Ye, Brandon and Alyakin, Anton and Markert, John E. and Hsieh, Aaron and Mańkowski, Michał and Oermann, Eric K.},
  year={2026}, eprint={2610.08585}, archivePrefix={arXiv}, primaryClass={cs.CL},
  url={https://arxiv.org/abs/2610.08585}
}
```