---
pretty_name: NPCBench
license: cc-by-4.0
configs:
- config_name: L1
  data_files:
  - split: test
    path: L1/questions.jsonl
- config_name: L2
  data_files:
  - split: test
    path: L2/questions.jsonl
- config_name: L3-Path
  data_files:
  - split: test
    path: L3/Path.jsonl
- config_name: L3-Seq
  data_files:
  - split: test
    path: L3/Seq.jsonl
- config_name: L3-Struct
  data_files:
  - split: test
    path: L3/Struct.jsonl
- config_name: L3-T100-anatomical
  data_files:
  - split: test
    path: L3/T-Stg.jsonl
- config_name: L4
  data_files:
  - split: test
    path: L4/questions.jsonl
- config_name: L5
  data_files:
  - split: test
    path: L5/questions.jsonl
- config_name: O1
  data_files:
  - split: test
    path: O1/questions.jsonl
- config_name: L3-T400
  data_files:
  - split: test
    path: L3/T400.jsonl
- config_name: L3-N400
  data_files:
  - split: test
    path: L3/N400.jsonl
language:
- en
- zh
tags:
- medical
- benchmark
- multimodal
- nasopharyngeal-carcinoma
---
# NPCBench

NPCBench evaluates guideline knowledge, image interpretation, care decisions,
and specialist consultation in nasopharyngeal carcinoma.

[Paper](https://openreview.net/forum?id=AVTJEREEah) · [Evaluation code](https://github.com/pidneuralode/NPCBench)

## Tasks

- **L1 and L2** contain 250 guideline questions and 185 application questions.
- **L3** contains pathology, MRI sequence recognition, structure assessment, and T400/N400 staging. T400 uses structure cards for 100 patients and sequence groups for 300. N400 uses sequence groups from a disjoint cohort.
- **L4 and L5** contain 165 planning questions and 305 longitudinal decisions, respectively.
- **L6** covers 100 patients and 417 scored tasks (100 pathology assessments, 100 MRI T assessments, and 217 care decisions). The archive records six excluded treatment questions and their reasons.
- **O1** contains 103 consultations from 42 cases and 1,726 scoring criteria.

## Download

```sh
hf download pengkaiScience/NPCBench L2/questions.jsonl --repo-type dataset --local-dir data
```

For image tasks, see the [L3 data guide](L3/README.md) and the
[CarePath guide](L6/full100_inputs_scored217/README.md). They give archive paths
and the information needed to match images, prompts, and labels.
[Supplementary data](supplementary/README.md) cover input controls, external MRI,
cine, and physician calibration.

## Use

The release uses de-identified clinical inputs. L4 retains dates needed to apply
the relevant guideline edition. Longitudinal tasks use patient-relative days
where applicable. Ethics approval is documented under IRB B2026-465-01.
NPCBench is intended for research evaluation. Retrospective, disease-specific
sampling and differences in image formats affect how scores should be interpreted.

Original data and annotations use [CC BY 4.0](LICENSE-DATA.md). Third-party
materials retain their terms. Please cite the paper when using the dataset.