---
license: cc-by-4.0
language:
- en
task_categories:
- question-answering
- text-generation
task_ids:
- closed-domain-qa
- abstractive-qa
- dialogue-generation
size_categories:
- 10K<n<100K
tags:
- medical
- oncology
- benchmark
- tumor-board
- clinical-reasoning
- evaluation
- cancer
- clinical
- healthcare
- health
- biology
- reasoning
- llm-evaluation
- llm-as-a-judge
pretty_name: OpenTumorBoard
configs:
- config_name: board_simulation
  default: true
  data_files:
  - split: train
    path: data/task1_train.jsonl.gz
  - split: validation
    path: data/task1_validation.jsonl.gz
  - split: test
    path: data/task1_test.jsonl.gz
- config_name: specialist_turn
  data_files:
  - split: train
    path: data/task2_train.jsonl.gz
  - split: validation
    path: data/task2_validation.jsonl.gz
  - split: test
    path: data/task2_test.jsonl.gz
---

# OpenTumorBoard

**[Paper](https://arxiv.org/abs/2609.32810) · [Code](https://github.com/AnqiLi24/OpenTumorBoard) · [Leaderboard](https://huggingface.co/spaces/al1219/OpenTumorBoard-Leaderboard)**

OpenTumorBoard evaluates language models on cases discussed at real multidisciplinary tumor boards. This release contains derived benchmark annotations from **611 patient cases across 219 publicly recorded meetings**, with **16,215 specialist-turn question–answer pairs**.

This is a caption-only distribution of the current **v1.0** benchmark data. It preserves the source benchmark records and recording-level splits. The benchmark itself supports image-based evaluation, but **slide image files are not included in this release**.

## Tasks and splits

| Configuration | Unit | Train | Validation | Test |
|---|---|---:|---:|---:|
| `board_simulation` | Patient case | 366 | 61 | 184 |
| `specialist_turn` | Question–answer pair | 9,731 | 1,640 | 4,844 |
| Recording membership | Recording | 131 | 22 | 66 |

**Board Simulation** provides a case summary and slide captions for generating a multidisciplinary discussion and a final conclusion. The reference target in this distribution is the board-level conclusion.

**Specialist Turn** provides a case summary, slide captions, a question, and a target specialist role. The reference answer is derived from the recorded specialist response. Each item has one of nine question types.

Recordings are assigned to exactly one split. A recording does not appear in both training and validation or test data.

## Loading the data

The two configurations load separately with the Hugging Face `datasets` library:

```python
from datasets import load_dataset

repo_id = "al1219/OpenTumorBoard"
board = load_dataset(repo_id, "board_simulation")
specialist = load_dataset(repo_id, "specialist_turn")

print(board["test"][0]["reference_conclusion"])
print(specialist["test"][0]["qa_type"])
```

The data files use gzip-compressed JSONL: each decompressed line is one JSON object. The `slides` field contains **text captions**, not image pixels.

## Fields

Both task configurations include:

- `schema_version`, `benchmark_version`: source record format and benchmark version.
- `video_uid`, `video_dir`, `case_id`: recording and case identifiers. Recording titles are retained as provenance. `video_uid` uses sequential identifiers such as `video_0001`; `simulation_id` and `qa_id` similarly use `sim_000001` and `qa_000001`.
- `case_summary`, `slides`: case summary and slide-caption text.
- `source_task`: source task provenance.

Board Simulation also includes `task_name`, `simulation_id`, `case_start_sec`, `case_end_sec`, and `reference_conclusion`.

Specialist Turn also includes `qa_id`, `qa_index`, `qa_type`, `qa_type_candidates`, `utterance_id`, `target_specialist_role`, `question`, `reference_answer`, and `source_aliases`. Some records carry optional `split` or `question_screening` metadata inherited from the source release.

The nine `qa_type` values are `agreement_or_support`, `clarification_question`, `clinical_trial_suggestion`, `eligibility_assessment`, `evidence_discussion`, `findings_interpretation`, `next_action_suggestion`, `treatment_recommendation`, and `uncertainty`.

## Additional files

| File | Contents |
|---|---|
| `data/split.json` | Recording membership, source YouTube URLs, and split statistics |
| `data/slides.jsonl.gz` | 4,923 slide-source records with URLs and time intervals |
| `data/slides_index.jsonl.gz` | 16,826 case/question records mapping to relative slide paths |
| `data/test_inputs/board_simulation.caption.jsonl.gz` | 184 prepared caption-only test inputs |
| `data/test_inputs/specialist_turn.caption.jsonl.gz` | 4,844 prepared caption-only test inputs |

Slide paths in the metadata identify media associated with the source recordings. They do not contain image pixels and do not imply that the corresponding files are bundled here. The prepared test inputs contain prompts and input metadata; reference answers are provided in the task data files.

## Media and evaluation scope

This release includes derived text annotations, slide captions, source links, timestamps, and provenance metadata. It does **not** include slide images, original videos, audio, raw verbatim transcripts, or complete discussion trajectories used as training targets.

Image-enabled evaluation requires separately available source media and the relevant preprocessing steps. Caption-only runs and image-enabled runs use different inputs and should be identified accordingly when reporting results.

## Version and provenance

The ten compressed JSONL files preserve the source release's benchmark content, with hash metadata fields removed and recording, simulation, and question identifiers replaced by sequential identifiers. The same identifiers are used across task records, prepared inputs, the slide index, and split metadata. In `split.json`, question counts, question-type counts, and achieved question-count ratios have been recalculated from the released post-screening files. The source split metadata documented its statistics as pre-screening counts. Recording membership, split assignments, source URLs, seed, and clinical content have not changed.

## Limitations and rights

The data originate exclusively from publicly available tumor board recordings and may not represent all institutions, patient populations, or clinical workflows. Derived annotations can contain extraction or annotation errors. This benchmark is intended for research evaluation and is not a substitute for clinical decision-making.

The original recordings, presentation slides, and verbatim speech remain subject to the rights and terms of their respective owners. This distribution does not redistribute that source media or grant rights over it. The benchmark annotations in this release are licensed under CC BY 4.0. The license covers the annotations only and does not extend to the source media above.

The source paper states that no independent patient-identifier screening was performed beyond the privacy protections applied in the publicly released recordings. This release should not be interpreted as an independently verified de-identification of the source material.

## Citation

If you use OpenTumorBoard in your research, please cite the paper:

```bibtex
@misc{li2026opentumorboard,
  title         = {{OpenTumorBoard}: A Real-World Benchmark of Multidisciplinary Tumor Board Discussion Trajectories},
  author        = {Anqi Li and Zhixuan Ge and Yixuan Duan and Jiarong Qian and Chi-Yu Chen and MingYu Lu and Huan-Yu Hsu and Yu Gu and Yue Guo and Sheng Wang and Wei Qiu and Hanwen Xu},
  year          = {2026},
  eprint        = {2609.32810},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  doi           = {10.48550/arXiv.2609.32810},
  url           = {https://arxiv.org/abs/2609.32810}
}
```