Database Open Access

VitalDB - AnesLLM: a dataset and benchmark based on real intraoperative anaesthetic decisions

ORCID iD for Ramses Marrero Garcia,  ,  ,  ,  ORCID iD for Hyung-Chul Lee

Published Nov. 20, 2026 · Version 1.0
When using this resource, please cite:

Marrero Garcia, R., Peñafiel Macías, P. A., Cayuela Perez, R., Borrat Frigola, X., & Lee, H. (2026). VitalDB - AnesLLM: a dataset and benchmark based on real intraoperative anaesthetic decisions (version 1.0). PhysioNet. RRID:SCR_007345. https://doi.org/10.13026/yexv-rf10

Please include the standard citation for PhysioNet:

Pollard, T., Moody, B. E., Lehman, L., Gow, B., Fernandes, C., Xie, C., Johnson, A., Mark, R. G., & Heldt, T. (2026). PhysioNet as a global platform for biomedical research. Nature Health. https://doi.org/10.1038/s44360-026-00096-z. Available from: https://pmc.ncbi.nlm.nih.gov/articles/PMC13641116/

Abstract

AnesLLM is a dataset of real intraoperative anaesthetic decisions derived from VitalDB, an open database of high-resolution recordings from non-cardiac surgery at Seoul National University Hospital. It contains 84 813 decision windows from 3168 anaesthetic cases. Each window is anchored at a decision point and contains only information recorded before it: patient and surgical data, the last preoperative value of 34 laboratory tests, 10 min of vital signs as time series, drug concentrations and ventilatory parameters at the decision point, and a 20-min history of previous actions (280 fields in total). The label is the action actually taken by the anaesthetist, detected automatically from infusion-pump changes: increase or reduce hypnotic, increase or reduce opioid, or no action. Because more than one action is often acceptable, each window also includes the two actions most frequently chosen by the cohort’s anaesthetists in similar clinical states (consensus actions), with their estimated probabilities; vasopressor may appear as an additional, investigator-defined consensus option. Cases are assigned to training (2994), development (25) and test (149) partitions at case level, and the test partition is stratified by case difficulty. The release includes a data dictionary, a partition manifest, a datasheet, machine-readable audit reports (integrity, information leakage, plausibility, drift, missingness and k-anonymity) and command-line utilities to filter windows, convert them to tables or text clinical cases, and verify the integrity of the download. AnesLLM is intended to evaluate any artificial intelligence system on choosing the next action at a specific moment of a real anaesthetic.


Background

Intraoperative anaesthetic care involves a continuous stream of titration decisions. During maintenance, the anaesthetist repeatedly reassesses the depth of anaesthesia, the haemodynamic state and the drug concentrations, and either adjusts the hypnotic or the opioid infusion or leaves the regimen unchanged. Practice in this setting is heterogeneous: different clinicians and institutions handle comparable patients in different ways [@naik2021], so for a given moment there is often more than one reasonable course of action.

Interest in artificial intelligence systems that could assist with these decisions is growing, but the resources used to evaluate them have important gaps. Most evaluations of language models in health care rely on examination-style questions or constructed vignettes rather than real patient data [@bedi2025], and such benchmarks have been criticised for assuming a single correct answer and for measuring knowledge rather than performance in clinical work [@raji2025]. When models are asked to work with real clinical records across several steps, their performance drops considerably [@hager2024]. To our knowledge, no public dataset pairs real intraoperative titration decisions with the information that was available to the anaesthetist just before each one.

VitalDB provides the raw material to build such a resource: high-resolution intraoperative recordings of vital signs, drug infusions and ventilation, together with clinical data, from surgical patients at Seoul National University Hospital, openly available on PhysioNet [@lee2022, @lee2022b]. AnesLLM is derived from VitalDB. It reorganises these recordings into decision windows, each containing the pre-decision state of the patient and the action the anaesthetist actually took. Given the variability of practice, each window also carries a plural reference: the two actions most frequently chosen by the cohort’s anaesthetists in comparable clinical states. Users can therefore choose how strictly to score a system.

The resource is designed to be model-agnostic. The same windows can be used to evaluate language models, through a deterministic conversion to text clinical cases, as well as tabular classifiers, time-series models, reinforcement-learning policies or rule-based controllers.


Methods

Source data

AnesLLM was derived from VitalDB version 1.0.0 [@lee2022, @lee2022b], which contains 6388 anaesthetics for non-cardiac surgery (general, thoracic, urological and gynaecological; elective and emergency) performed at Seoul National University Hospital between August 2016 and June 2017. VitalDB records continuous physiological signals, drug infusion data, ventilator parameters and clinical information captured with Vital Recorder [@lee2018]. Preoperative laboratory results were obtained from the VitalDB laboratory data. Pulse pressure variation and stroke volume variation were not used because they contained frequent artefactual values above 100%.

Decision windows

The unit of the dataset is the decision window. Each window is anchored at an instant t=0, at which a decision is observed. Only information recorded before t=0 is used as input. This input combines time-stamped information from the minutes preceding the decision with values describing the patient at t=0 or before surgery:

  • vital signs over the previous 10 min as time series (heart rate, mean, systolic and diastolic arterial pressure, SpO₂, EtCO₂ and bispectral index), with their current values, extremes, means and trends;
  • the actions and events of the previous 20 min, each with its timestamp;
  • at t=0, the effect-site concentrations of propofol and remifentanil, the end-tidal sevoflurane fraction and the minimum alveolar concentration;
  • at t=0, the ventilator settings (FiO₂, PEEP, tidal volume and respiratory rate);
  • patient and surgical characteristics without timestamps (age, sex, weight, height, body mass index, ASA physical status, procedure, elective or emergency status, diagnosis, department and type of monitoring);
  • the last preoperative value of 34 laboratory tests and the time at which each was recorded.

The 10-min and 20-min horizons were chosen to describe the current haemodynamic state and its recent trajectory while remaining close to the moment of the decision. Windows are stored in a semi-structured format (one JSON object per window), so that they can be used directly by any type of model.

Detection of actions

Actions were detected automatically from the infusion-pump records, not from annotations entered by clinicians. The thresholds were fixed before extraction: a change of at least 0.2 µg/ml in the propofol target concentration, or of at least 0.3 ng/ml in the remifentanil plasma target concentration. Events had to be separated by at least 120 s, so that a single adjustment does not produce several near-identical windows. As a consequence, rapid cascades of adjustments (for example, a propofol change followed shortly by a remifentanil change) are not represented.

The action actually taken at t=0 (result_real) belongs to one of five classes: increase_hypnotic, reduce_hypnotic, increase_opioid, reduce_opioid or no_action. Hypnotic actions refer to propofol only; sevoflurane changes and ventilator changes were not treated as actions.

Vasopressor administration is not reliably recorded in VitalDB, so it could not be detected as an actual action. To allow evaluated systems to propose it, the investigators marked vasopressor as an additional option in windows meeting a conservative haemodynamic criterion: mean arterial pressure below 65 mm Hg sustained for at least 30 s, with a stable propofol target concentration over the previous 180 s. The criterion favours precision over coverage.

Selection of decision points

Two kinds of windows were generated: windows anchored at a detected infusion change, and windows anchored at moments with no change. In both cases the action at t=0 had to be canonical, that is, a single identifiable decision rather than part of a sequence of adjustments. The same 5-min rule was applied to every class:

  • Action windows: an infusion change was kept only if no other action occurred within the following 5 min.
  • No-action windows: points were not drawn from every recorded instant but defined by the rule itself. A point was included when no action occurred at t=0 or in the following 5 min, and all points meeting this rule were kept, without random sampling.

The 5-min horizon reflects the interval within which the anaesthetist is expected to review the monitors systematically. No-action windows therefore represent an assumed decision not to intervene rather than mere periods without change; this is a modelling assumption. Information after t=0 was used only to decide which points enter the dataset, never as input.

Consensus actions

For each window, two consensus actions were estimated: the first (result_1) and second (result_aux) most likely responses of the cohort’s anaesthetists in similar clinical states, each with its estimated probability (ratio). The label set included the five actual-action classes plus vasopressor in the marked windows. No language model was involved.

The consensus was generated with LightGBM [@ke2017]. Among the strategies evaluated, it showed the highest agreement with the actions actually taken, and it handles missing values natively, without imputation. Probabilities were obtained out of fold, using three folds grouped by case, so that no window was ever predicted by a model trained on windows from the same case [@kapoor2023]. The comparison of strategies is reported in the accompanying publication.

Case difficulty

Difficulty was defined per case, so that all windows of a case share the same level. The confidence of a window is the sum of the probabilities of its two consensus actions, and a case’s score is the mean confidence of its windows. Cases above the 75th percentile were classified as easy, those below the 25th percentile as hard, and the rest as medium. Difficulty reflects how predictable the anaesthetists’ behaviour in a case is for a learning system, not its clinical severity.

Partitioning

Windows from the same case can share up to 20 min of input and are therefore not independent. For this reason, the dataset was partitioned by case rather than by window, and each case was assigned to a single partition:

  • Development: the cases with the highest proportion of missing data, retaining their artefactual values, to test robustness to incomplete inputs.
  • Test: cases stratified by difficulty (25% easy, 50% medium, 25% hard) and free of physiologically impossible values; any candidate case containing such values was moved to training.
  • Training: all remaining cases.

The assignment is deterministic and recorded in a partition manifest.

Quality control

The dataset was audited for:

  • integrity: partition exclusivity, unique window identifiers, ordered timestamps, and valid labels and probabilities;
  • information leakage: no label-derived or future variable in the input [@kapoor2023];
  • numerical plausibility;
  • distribution shift between training and test;
  • missingness;
  • window-level k-anonymity.

In addition, three anaesthetists from outside the source hospital, trained in different countries, independently reviewed a random sample of windows. They were blinded to the origin of each candidate action and judged whether the actual and consensus actions were appropriate. The review was not used to select the consensus strategy. The audit results are distributed as reports with the dataset.


Data Description

Files

The dataset is organised as follows:

  • data/train/, data/dev/, data/test/: one JSONL file per case (case<id>.jsonl), with one JSON object per window and line.
  • reports/split_manifest.csv: partition and difficulty level of each case.
  • reports/data_dictionary.csv: definition of the 292 variables (280 input fields, 7 vital-sign series and 5 output fields).
  • reports/: machine-readable audit reports (integrity, output consistency, information leakage, class balance, numerical plausibility, drift, missingness, coverage per variable and k-anonymity).
  • reports/dataset_sha256.json: SHA-256 fingerprint of the release.
  • docs/: dataset description, design notes, audit report, evaluation guide and datasheet.
  • scripts/: command-line utilities for summarising, filtering, sampling, converting and verifying the data (see Usage Notes).

Record structure

Each window is a JSON object with four top-level fields:

Field Type Content
window_id string Unique window identifier
case_id string VitalDB case identifier
input object 280 pre-decision fields and 7 vital-sign series
output object result_real, result_1 and result_aux

The input fields are grouped as:

  • patient and surgical data (72);
  • laboratory (68);
  • trends (56);
  • signals (42);
  • metadata (42);
  • vital-sign series (7).

The output contains:

  • result_real: the action actually taken (five classes);
  • result_1 and result_aux: the first and second consensus actions (six classes, including vasopressor), each with its probability (ratio).

The case identifiers (pt_caseid, pt_subjectid) are the synthetic identifiers used in VitalDB. They are retained so that windows can be traced back to the source database. Sex is the variable recorded in VitalDB; no information on gender is available.

Composition

The dataset contains 84 813 windows from 3168 cases. The remaining 3220 VitalDB cases had no window meeting the selection rule. The median number of windows per case is 23 (range 3–130).

Training Development Test
Cases 2994 25 149
Windows 80 933 423 3457
Windows per case, mean (SD) 27.0 (16.1) 16.9 (8.2) 23.2 (15.0)
Age, years, mean (SD) [range] 57.8 (14.4) [6–92] 56.1 (11.2) [39–75] 57.4 (14.4) [23–88]
Women, n (%) 1484 (49.6) 15 (60.0) 74 (49.7)
Duration of anaesthesia, h, mean (SD) 3.40 (1.91) 2.06 (0.70) 2.95 (1.90)
Invasive arterial pressure, % of windows 73.7 10.9 54.4

All ten paediatric cases are in the training partition.

Cases were classified by difficulty as 792 easy, 1584 medium and 792 hard. The case-level thresholds were a mean window confidence of 0.8318 (75th percentile) and 0.7772 (25th percentile).

Labels

Action Actual action, n (%) First consensus, n (%) Second consensus, n (%)
No action 23 638 (27.9) 33 583 (39.6) 7301 (8.6)
Reduce opioid 23 532 (27.7) 21 795 (25.7) 29 323 (34.6)
Increase opioid 20 630 (24.3) 19 420 (22.9) 27 355 (32.3)
Reduce hypnotic 10 231 (12.1) 5949 (7.0) 13 959 (16.5)
Increase hypnotic 6782 (8.0) 3669 (4.3) 6003 (7.1)
Vasopressor — 397 (0.5) 872 (1.0)

The class distribution reflects the decision points that met the selection rule, without resampling or weighting. It does not reflect the proportion of anaesthetic time without intervention, which is much larger.

The first and second consensus actions are always different. The first consensus action has a mean probability of 57.1% (median 55.5%, interquartile range 44.3–68.3%) and the second 23.0% (median 22.9%, interquartile range 15.5–30.0%). Together they cover on average about 80% of the cohort’s estimated behaviour.

Vasopressor appears as a consensus option in 1269 windows (1.5%): 1236 in training, 33 in test and none in development. It never appears as an actual action.

Agreement between consensus and actual actions

The actual action matched the first consensus action in 55.8% of windows and either consensus action in 80.4%. Agreement decreased with case difficulty:

Cases First consensus Either consensus
Easy 61.5% 86.6%
Medium 56.5% 80.9%
Hard 50.1% 74.7%
Test partition 56.8% 81.3%
Development partition 49.4% 73.8%

External review

The three external reviewers assessed 298 windows from 272 cases, making 907 judgements. Only 13 of these windows are in the test partition.

  • Inter-rater agreement: moderate (Fleiss’ κ 0.515; pairwise Cohen’s κ 0.467–0.609).
  • Actual action: approved by majority in 24.5% of windows (95% CI 19.3–29.4%).
  • First consensus action: approved by majority in 26.6% (21.9–31.8%).
  • Second consensus action: approved by majority in 54.5% (45.8–64.2%) of the windows in which it was assessed.

Confidence intervals were obtained by case-clustered bootstrap.

Data quality

  • Integrity: no case appears in more than one partition, window identifiers are unique and no case has decreasing timestamps. In 526 cases, two consecutive windows share the same timestamp.
  • Information leakage: none of the 21 variables derived from the labels, the consensus strategy or post-decision information is present in the input.
  • Numerical plausibility: 1730 windows (2.1% of training and development windows; 864 training cases and 1 development case) contain physiologically impossible haemodynamic values arising from VitalDB signal artefacts. They are retained and documented case by case rather than corrected. The test partition contains none. In addition, 27 creatinine values exceed 20 mg/dl and one minimum alveolar concentration value exceeds 3.
  • Missingness: the mean proportion of null fields per window is 0.116 (median 0.085, range 0.014–0.381): 0.115 in training, 0.288 in development and 0.133 in test. It is similar across easy, medium and hard cases. Invasive arterial pressure is null in windows monitored non-invasively.
  • Drift between training and test: the total variation distance between the actual-action distributions is 0.011 (population stability index [PSI] 0.017). Of 19 input variables compared, 14 have a PSI below 0.10 and five show a moderate shift (PSI 0.10–0.25): potassium, platelets, heart rate, age and haemoglobin. None exceeds 0.25.
  • Privacy: with sex, age, weight, height, body mass index and ASA physical status as quasi-identifiers, each window shares its profile with a median of 23 windows (minimum k=3). Windows from the same patient repeat the same profile.

Usage Notes

Reading the data

Each file in data/ is a JSONL file and can be read line by line with any standard JSON reader. Every line is one window. Variable definitions are given in reports/data_dictionary.csv. Windows can be grouped by case_id, and the partition and difficulty of each case can be obtained from reports/split_manifest.csv. After downloading, users can check that their copy matches the release with scripts/check_integrity.py and scripts/dataset_sha256.py.

Utilities

The scripts/ folder contains command-line tools that work directly on the JSONL files:

  • quick_stats.py: summary of cases, windows and labels;
  • filter.py: filtering by label, difficulty, sex, monitoring type or case;
  • sample.py: random, per-class or stratified sampling;
  • to_csv.py: conversion to a flat table for tabular models;
  • to_prompt.py: conversion of a window into a short clinical case in English, with an example prompt (prompt.txt);
  • check_integrity.py and dataset_sha256.py: integrity and fingerprint verification.

Input format: structured or text

The dataset is distributed in semi-structured form, and this is its primary format. Machine-learning models (tabular classifiers, time-series models, reinforcement-learning policies or rule-based controllers) can be evaluated directly on the structured input, or on its CSV conversion.

For language models, we recommend converting each window into a clinical case with to_prompt.py. This conversion is deterministic: it omits missing values and never includes identifiers. It is an option for text-based systems, not a required step. Users may design their own representation, provided it uses only the input fields.

Recommended evaluation protocol

  • Partitions: report results on the test partition. Use the development partition to tune prompts or hyperparameters, and the training partition for training.
  • Metrics: three accuracies are defined.
    • Plausible accuracy, the primary metric: match with either consensus action. It reflects that more than one action may be acceptable.
    • Consensus accuracy: match with the first consensus action.
    • Strict accuracy: match with the action actually taken.
  • Trivial floors: compute the same metrics for strategies that require no clinical judgement: never act, always predict the most frequent class, and repeat the previous action of the case. With imbalanced classes and a plural reference, absolute accuracies cannot be interpreted without these floors.
  • Human reference: in the test partition, the anaesthetists’ own actions reach 0.568 consensus accuracy and 0.813 plausible accuracy. This is not a ceiling, since an anaesthetist with full information does not reach 1. It marks the level of human practice measured with the same instrument. A system is best described by its distance from this reference and from the trivial floors.
  • Uncertainty: windows from the same case are not independent. Confidence intervals should be obtained by case-clustered bootstrap, and comparisons between systems should use methods that respect this clustering [@field2007].
  • Safety: physiological contraindication rules can be applied to the proposed actions using the same input variables. They should also be applied to the actual actions, as a human reference.

Windows with vasopressor

Vasopressor is an investigator-defined consensus option, not a recorded action. Users who do not accept the criterion used to mark it should exclude the windows in which either consensus action is vasopressor. These windows must also be excluded when the evaluated system cannot propose a vasopressor.

Other uses

  • Training: the training partition (more than 80 000 windows) can be used to fine-tune or adapt models, study retrieval of similar cases, train reinforcement-learning policies or compare classical controllers.
  • Robustness: the development partition, which concentrates incomplete and artefactual cases, can be used to study robustness to missing data.
  • Difficulty levels: these allow stratified evaluation and curriculum learning.

Limitations

  • Centre and period: a single centre, 2016–2017, and mainly total intravenous anaesthesia with propofol and remifentanil.
  • Action space: only the direction of change is captured, not its magnitude. Sevoflurane and ventilator changes are not actions, and vasopressor is only an investigator-defined option.
  • Window selection: the selection rule excludes periods of rapid, chained adjustments. No-action windows rely on the assumption that no change within 5 min reflects a decision not to intervene.
  • Reference: the consensus estimates the cohort’s most common behaviour, not the best possible decision, and neither the consensus nor the actual action is linked to patient outcomes. The dataset measures agreement with practice, not clinical correctness.
  • Label noise: actual actions are defined by infusion thresholds and may include technical adjustments without clinical intent, such as syringe changes.

Inappropriate uses

  • Using post-decision information or the output fields as input.
  • Evaluating on the training partition, or mixing windows from the same case between training and evaluation.
  • Using the dataset to guide decisions on patients, or to claim that a system is clinically safe without prospective validation.
  • Attempting to re-identify patients.

Release Notes

Version 1.0

Initial release of this dataset.


Ethics

This dataset was developed within the AnesLLM-2026 study, approved by the Research Ethics Committee (CEIm) of Hospital Clínic de Barcelona (registration HCB/2026/0831; favourable opinion of 17 August 2026). The approved study covers the extraction and preparation of the source data and the construction and validation of the derived dataset as an in silico study. The committee waived informed consent because the source data are public and were anonymised at source.

The source data are VitalDB version 1.0.0 [@lee2022, @lee2022b]. VitalDB was collected with the approval of the Institutional Review Board of Seoul National University Hospital (H-1408-101-605) and registered at ClinicalTrials.gov (NCT02914444). Written informed consent was waived because of the anonymity of the data, and the dataset was de-identified before its public release. VitalDB is distributed under the Creative Commons Attribution 4.0 International (CC BY 4.0) licence, which permits redistribution of derived works with attribution and indication of changes. AnesLLM contains no information beyond that already present in VitalDB. It retains only VitalDB’s synthetic case and subject identifiers, and it is released under the same licence.


Acknowledgements

We thank Seoul National University Hospital and all of its anaesthesiologists, whose daily clinical work is recorded in VitalDB and made this dataset possible. The decisions captured in AnesLLM are theirs. We are also grateful to the VitalDB team for making these high-resolution intraoperative recordings openly available to the research community. We also thank the three external anaesthesiologists who independently reviewed a sample of windows.


Conflicts of Interest

The authors have no conflicts of interest to declare.


References

  1. @naik2021 Naik, B. I., Kuck, K., Saager, L., Kheterpal, S., Domino, K. B., Posner, K. L., Sinha, A., Stuart, A., Brummett, C. M., Durieux, M. E., Vaughn, M. T., Pace, N. L., & . (2021). Practice Patterns and Variability in Intraoperative Opioid Utilization: A Report From the Multicenter Perioperative Outcomes Group. Anesthesia &Amp; Analgesia, 134(1), 8–17. https://doi.org/10.1213/ANE.0000000000005663
  2. @bedi2025 Bedi, S., Liu, Y., Orr-Ewing, L., Dash, D., Koyejo, S., Callahan, A., Fries, J. A., Wornow, M., Swaminathan, A., Lehmann, L. S., Hong, H. J., Kashyap, M., Chaurasia, A. R., Shah, N. R., Singh, K., Tazbaz, T., Milstein, A., Pfeffer, M. A., & Shah, N. H. (2025). Testing and Evaluation of Health Care Applications of Large Language Models. Jama, 333(4), 319. https://doi.org/10.1001/jama.2024.21700
  3. @raji2025 Raji, I. D., Daneshjou, R., & Alsentzer, E. (2025). It’s Time to Bench the Medical Exam Benchmark. Nejm Ai, 2(2). https://doi.org/10.1056/AIe2401235
  4. @hager2024 Hager, P., Jungmann, F., Holland, R., Bhagat, K., Hubrecht, I., Knauer, M., Vielhauer, J., Makowski, M., Braren, R., Kaissis, G., & Rueckert, D. (2024). Evaluation and mitigation of the limitations of large language models in clinical decision-making. Nature Medicine, 30(9), 2613–2622. https://doi.org/10.1038/s41591-024-03097-1
  5. @lee2022 Lee, H.-C., Park, Y., Yoon, S. B., Yang, S. M., Park, D., & Jung, C.-W. (2022). VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients. Scientific Data, 9(1). https://doi.org/10.1038/s41597-022-01411-5
  6. @lee2022b Lee, H.-C., & Jung, C.-W. (2022). VitalDB, a high-fidelity multi-parameter vital signs database in surgical patients (Version 1.0.0) [Dataset]. PhysioNet. https://doi.org/10.13026/czw8-9p62
  7. @lee2018 Lee, H.-C., & Jung, C.-W. (2018). Vital Recorder—a free research tool for automatic recording of high-resolution time-synchronised physiological data from multiple anaesthesia devices. Scientific Reports, 8(1). https://doi.org/10.1038/s41598-018-20062-4
  8. @ke2017 Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., & Liu, T.-Y. (2017). LightGBM: A highly efficient gradient boosting decision tree. In Advances in Neural Information Processing Systems 30 (pp. 3146–3154). Curran Associates.
  9. @kapoor2023 Kapoor, S., & Narayanan, A. (2023). Leakage and the reproducibility crisis in machine-learning-based science. Patterns, 4(9), 100804. https://doi.org/10.1016/j.patter.2023.100804
  10. @field2007 Field, C. A., & Welsh, A. H. (2007). Bootstrapping Clustered Data. Journal of the Royal Statistical Society Series B: Statistical Methodology, 69(3), 369–390. https://doi.org/10.1111/j.1467-9868.2007.00593.x

Files

Total uncompressed size: 950.3 MB.

Access the files
Folder Navigation: <base>
Name Size Modified
data
docs
reports
scripts
LICENSE.txt (download) 14.5 KB 2026-11-18
README.md (download) 4.9 KB 2026-10-21
SHA256SUMS.txt (download) 284.9 KB 2026-11-18