You need to agree to share your contact information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
The RadKev-27B weights are released for non-commercial research use only. Several training sources are licensed CC BY-NC-SA 4.0 (Eurorad, CT-RATE), and the model is not a medical device. By requesting access you agree to use the weights for non-commercial research only, not to use them for clinical decisions, and to share derivatives under the same terms.
Log in or Sign Up to review the conditions and access this model content.
RadKev-27B
RadKev-27B is an open-weight decision model for radiology. Given a state (a report, a case or a clinical indication) and one or
more typed questions, it returns, in a single forward pass and without generating text, a probability for every admissible
answer. It was obtained by continued training of Kev-27B (revision 01b8199) on 36,109
records from public radiology datasets: a rank-16 LoRA adapter on every linear projection of the frozen
Qwen/Qwen3.8-27B backbone, together with Kev's pointer head. Temperature fitted on the development split:
T = 1.35.
Code and documentation · Model card (full) · Released data · RadKev-9B
Results
Radiology benchmark of 14,142 held-out questions in fifteen tasks (report reading, case diagnosis and classification, radiological knowledge). Accuracy (%), unweighted mean of the task accuracies; 95% confidence intervals of paired differences from a source-stratified cluster bootstrap.
| System | Benchmark task mean (15 tasks) | Human-labeled tasks (13, every system) |
|---|---|---|
| RadKev-27B | 80.8 | 81.1 |
| RadKev-9B | 77.0 | 77.0 |
| Kev-27B | 75.9 | 77.3 |
| Kev-9B | 70.1 | 72.0 |
| Jev (jev-1.13.0) | 77.8 | 79.1 |
| OpenAI Decisions (gpt-6-luna) | 76.0 | 79.4 |
| Qwen3.8-27B | – | 78.1 |
RadKev-27B exceeded Kev-27B by 4.9 percentage points (95% CI 3.8 to 6.2), and was more accurate than Jev (3.0 points, 1.7 to 4.4), the OpenAI Decisions API (4.8 points, 3.2 to 6.4) and Qwen3.8-27B with reasoning disabled (4.7 points, 3.4 to 6.0) and, on a sample of 1,675 questions, enabled (2.8 points, 1.0 to 4.6), at 132-fold lower latency than reasoning (128 ms vs 16.9 s per question). At a 5% error rate it could answer 90.5% of the benchmark questions (Kev-27B 76.4%).
Usage
The checkpoint loads like any Kev checkpoint with Kev (the repository's scripts/setup.sh
installs Kev at the pinned commit):
python -m kev.serve --run ramu9703/radkev-27b --port 8009 # POST /v1/systemone
from radkev.predict import Predictor
model = Predictor("ramu9703/radkev-27b")
out = model({
"state": "FINDINGS: Small left pleural effusion. No pneumothorax.",
"questions": {"ptx": {"type": "noul", "instructions": "Does this report describe pneumothorax?"}},
})
out["answers"]["ptx"]["noul"] # P(yes)
Questions are of three types: noul (yes/no), choice (named options) and score (ordered levels). RadKev-27B requires approximately 55 GB in bfloat16: one 80 GB GPU, or two 48 GB GPUs with KEV_DEVICE_MAP=auto KEV_DTYPE=bf16 (the repository's two-GPU patch).
RadCases panel question. The reported results apply a prior correction on the RadCases ACR panel question: each option's
probability is divided by its add-one-smoothed frequency among the training-split panel answers and renormalized, because the
residual option ("no ACR Appropriateness Criteria topic applies") was the most frequent training answer (34%). The raw outputs of
these weights are uncorrected; the frequencies are in the released data (outputs/radcases_panel_prior.json).
Intended use and limitations
Research only. The model is not a medical device, has not been prospectively validated and must not be used for clinical decisions. It reads English text (reports, case descriptions, clinical indications), never images. The training and test data are public datasets and teaching collections; on an external test of radiologist-annotated reports from another institution, RadKev-27B was less accurate than Kev-27B on definite findings, and every decision model predominantly answered "present" for hedged findings. The model can only choose among the answers it is offered, so answer spaces should be verified to contain the correct answer, and the temperature and automation thresholds should be set on locally labeled cases.
License
Weights: CC BY-NC-SA 4.0, non-commercial research use (several training sources carry that license). Kev code: Apache-2.0. Backbone: see the Qwen/Qwen3.8-27B license.
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
- 2
Model tree for ramu9703/radkev-27b
Base model
Qwen/Qwen3.8-27B