You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These files accompany the RadKev manuscript. Records are distributed under the license of their source (Eurorad: CC BY-NC-SA 4.0; MedMCQA: Apache-2.0; MedQA: CC BY 4.0; MMLU, PubMedQA, MedXpertQA: MIT). Labels and model outputs are for non-commercial research use. By requesting access you agree to these terms and not to use the data for clinical decisions.
Log in or Sign Up to review the conditions and access this dataset content.
RadKev released data
Data released with RadKev: An Open-Weight Decision Model for Radiology (Udbhav Ram and Ran Zhang). Code, documentation and the
scripts that rebuild every record: https://github.com/udiram/RadKev (see docs/DATA.md). Models:
ramu9703/radkev-27b and ramu9703/radkev-9b.
Contents
| Path | Contents |
|---|---|
records/<source>/<split>.jsonl |
question records built from Eurorad, MedMCQA, MedQA, MMLU (medical subjects), PubMedQA and MedXpertQA, each under the license of its source |
labels/<source>/<split>.jsonl |
records released without text: identifier, group identifier (report, case or patient), SHA-256 of the state and each question with its ground-truth label. IU/Open-i, CT-RATE and RadCases carry the templated instructions and option names; ReXErr, RSNA-RadioQA and RadGraph-XL carry option identifiers and the label only |
labels/teacher/, labels/teacher_regenerated/ |
LLM-labeled questions on imaging orders, triage and follow-up, without the state: label and soft target. teacher holds the labels used for training; teacher_regenerated the labels regenerated after the MedGemma-27B-text read-out was corrected (Supplementary Note S2) |
labels/teacher_candidates.jsonl |
all candidate teacher questions with each teacher's answer distribution |
training/{train,dev}.jsonl |
the 36,109 training and 6,176 development records of RadKev-27B and RadKev-9B: source, group identifier, state hash and question identifiers, to be matched against records/ and labels/ by state hash |
test_index.jsonl |
every held-out test record: identifier, source, group identifier, state hash and, for each question, its task, its benchmark task (or null for questions outside the radiology benchmark) and whether its ground truth is human-labeled |
outputs/test/<system>.jsonl.gz |
per-question outputs of every system on the held-out test records: option identifiers, label, probabilities and benchmark task. Rows whose benchmark_task is not null form the radiology benchmark (14,142 questions; 14,047 for Qwen3.8-27B, which was not scored on the RadCases topic question) |
outputs/radcases_panel_prior.json |
training-split answer frequencies of the RadCases panel question, used for the prior correction of RadKev-27B and RadKev-9B |
outputs/reasoning/ |
Qwen3.8-27B with reasoning enabled on the reasoning sample (manuscript, Supplementary Note S4): probabilities, benchmark task and the read-out of each generation |
outputs/options_only/{full,blind}/ |
the options-only control: Eurorad and RSNA-RadioQA diagnosis questions with and without the case |
outputs/preread/{full,preread}/ |
subspecialty classification with and without the imaging findings |
answer_space/records.jsonl, outputs/answer_space/ |
the answer-space analysis: questions with each answer space (Eurorad with text; RSNA-RadioQA without text) and each system's probabilities and latencies |
outputs/external/<system>.jsonl.gz |
the RadGraph-XL external test (4,037 status questions; OpenAI Decisions and Jev as probabilities per record and question) |
outputs/transfer/<system>.jsonl.gz |
Kev's out-of-domain transfer suite, specialized and released models |
Systems. radkev-27b, radkev-9b; kev-27b, kev-9b (Kev-27B revision 01b8199, Kev-9B revision 2629c06);
qwen3.8-27b (option-letter probabilities, reasoning disabled, questions with at most 16 options); openai-decisions (OpenAI
Decisions API, gpt-6-luna) and jev (jev-1.13.0), whose probabilities are returned rounded to two decimals.
Identifiers. Test records are identified as in the manuscript's outputs: rad/<line> for the records shared with the IU,
Eurorad, CT-RATE and examination sources, and the record identifier for RadCases, ReXErr and RSNA-RadioQA. A question is
identified by its record identifier and question key.
RadCases panel question. The files contain the raw probabilities. The reported results of RadKev-27B and RadKev-9B divide each
panel option's probability by its add-one-smoothed training frequency (outputs/radcases_panel_prior.json) and renormalize.
Licenses and use
Records inherit the terms of their sources. IU/Open-i (CC BY-NC-ND 4.0) and the access-controlled sources (CT-RATE, CheXpert
Plus, ReXGradient-160K, ReXErr on MIMIC-CXR, RadGraph-XL) and RSNA-RadioQA are not released as text; rebuild them from the
original datasets with radkev.data and check the state hashes. Labels and model outputs are released for non-commercial research
use. Not a medical device; research use only.
Citation
@article{ram2026radkev,
title = {RadKev: An Open-Weight Decision Model for Radiology},
author = {Ram, Udbhav and Zhang, Ran},
year = {2026},
note = {Manuscript in preparation},
url = {https://github.com/udiram/RadKev}
}
- Downloads last month
- 31