PanCanBench

PanCanBench contains 282 de-identified, authentic pancreatic cancer patient and caregiver questions paired with 3,639 expert-designed, question-specific rubric items. It is designed to evaluate how well large language model (LLM) responses address the clinical and informational needs expressed in these questions.

Each question has a weighted rubric describing content that a response should include or avoid. The dataset is in English and is provided as a single evaluation (test) split containing all 282 questions.

Question collection

Source and selection

The Patient Services team at the Pancreatic Cancer Action Network (PanCAN) purposively selected questions from submissions to its Patient Services Help Line received during 2019–2025. Selection sought broad clinical coverage, including treatments, side effects, diet and nutrition, and supportive care. The team also sought a balance of question complexity, avoiding questions that were overly simple or highly specialized.

De-identification and review

PanCAN staff manually removed or modified identifying information in the questions. This process addressed patient names, geographic locations such as cities and states, specific family relationships, details of disease stage or subtype and treatment history, and references to healthcare providers or institutions. Four PanCAN team members independently reviewed the questions to check the de-identification before data sharing.

Only the question text was shared for this project. Patient contact information and additional patient metadata were not provided. The question identifiers in this dataset index benchmark entries; they are not patient identifiers.

Topic organization

Questions were initially grouped by AI into six categories, then manually reviewed and reassigned as needed:

• Treatment & Clinical Trials
• Diagnosis & Screening
• Supportive & Palliative Care
• Genetics & Risk
• Side Effects & Symptom Management
• Mechanistic / Research Topics

These categories describe the organization of the question collection. The JSON file contains question identifiers, question text, and rubric items; category labels are not included as a field in this file.

Rubric development

The human-curated rubrics were developed through four stages:

1. Independent drafting: Two oncology fellows independently drafted question-specific criteria for each assigned question.
2. Clarification: AI-assisted refinement helped clarify the criteria for consistent application, with expert review to preserve their clinical meaning.
3. Merging: AI merged semantically equivalent criteria from the two rubrics while retaining distinct criteria. The original two fellows reviewed the merged rubrics for missing criteria and incorrect additions.
4. Final expert review: A third oncology fellow reviewed and finalized each rubric, with authority to add, remove, modify, or merge criteria and adjust their weights.

The released rubrics contain between 2 and 30 items per question. Weights reflect the clinical and informational importance of each criterion. Positive items reward specified content; negative items penalize specified undesirable content or omissions.

Data structure

The file rubrics_all_questions_final_version.json is a JSON object with one top-level key, questions, containing a list of 282 records. Each record has the following fields:

Field	Type	Description
question_number	Integer	Unique benchmark question identifier, from 1 to 282.
question_text	String	De-identified patient or caregiver question.
rubric_items	List of objects	The question-specific scoring criteria.

Each object in rubric_items contains:

Field	Type	Description
item_number	Integer	Criterion identifier within this question. Use it together with question_number.
description	String	The criterion used to evaluate a response.
min_points	Integer	Lower scoring endpoint: 0 for a positive item, or a negative penalty weight.
max_points	Integer	Upper scoring endpoint: the positive reward weight, or 0 for a penalty item.

The two point fields encode binary outcomes, not a continuous partial-credit scale. For example, an item with min_points=0 and max_points=10 contributes either 0 or 10 points. An item with min_points=-5 and max_points=0 contributes either −5 or 0 points.

Load the dataset

Install the Hugging Face Datasets library:

pip install datasets

Load the evaluation split:

from datasets import load_dataset

dataset = load_dataset("YiminZ07/PanCanBench", split="test")

print(len(dataset))  # 282
print(dataset.column_names)
# ['question_number', 'question_text', 'rubric_items']
print(sum(len(items) for items in dataset["rubric_items"]))  # 3639

question = dataset[0]
print(question["question_text"])
print(question["rubric_items"][0])

The dataset configuration reads the top-level questions array so that each question becomes one row. The test split contains the complete benchmark; no training or validation split is supplied.

If you download the JSON file directly, load the nested array explicitly:

from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={"test": "rubrics_all_questions_final_version.json"},
    field="questions",
    split="test",
)

For reproducible evaluations, record the dataset commit and use the revision argument when loading from the Hub.

Scoring

Evaluate each response against the rubric attached to its question. For each positive item, award max_points if the response meets the criterion and 0 otherwise. For each negative item, apply min_points if the undesirable condition described in the item occurs and 0 otherwise. A penalty may refer to an omission, so the criterion's wording determines whether it applies.

For each question:

1. Sum all awarded positive points and applied negative penalties to obtain the signed total.
2. Sum the positive item weights (max_points) to obtain the maximum positive total.
3. Calculate the percentage score as:

question_score = 100 × max(0, signed_total) / maximum_positive_total

The score is floored at zero after penalties are applied. For example, 80 awarded positive points and a 10-point penalty, with 100 possible positive points, give a question score of 70.

A model's mean rubric score is the arithmetic mean of its question-level percentage scores, giving each evaluated question equal weight. Genuine zero scores are included. If a response was not generated, report that missing response and the number of evaluated questions separately; the study's scoring workflow excludes explicit missing-response records from the mean. Pooling raw points across questions produces a different statistic because questions have different total weights.

The rubrics support human or automated grading. Automated evaluations should also report the judge model and grading instructions, which can affect the decisions assigned to individual items.

Intended use and limitations

PanCanBench is intended for research evaluating responses to pancreatic cancer patient and caregiver questions. Its purposive sampling emphasizes clinical breadth and does not establish population-level representativeness. The English-language Help Line setting also limits generalization to other languages and care settings.

Rubric scores measure performance against the specified criteria. They do not establish that every statement in a response is factually correct or that the response is suitable for an individual patient's care. Clinical evidence and recommendations can change over time, and the released rubrics reflect their development context.

Related resources and citation

• Evaluation code and reproduction materials: PanCanBench on GitHub.
• Paper: Yimin Zhao et al. PanCanBench: A Comprehensive Benchmark for Evaluating Large Language Models in Pancreatic Oncology. 2026. arXiv:2603.01343.

The original preprint describes an earlier 3,130-item rubric set. This dataset contains the revised 3,639-item set. When citing an evaluation, identify the dataset revision as well as the paper.

License

The dataset is released under the MIT license.