Datasetpublic

kobzaond/RLVRAMBench

RLVRAMBench Which language-model training configurations can I use with the memory I have, and how much testing does that decision require? RLVRAMBench is a measurement dataset with open evaluation tasks for a specific language-model training system. It measures memory feasibility when response generation and reinforcement-learning updates share the same graphics processors. It provides measured outcomes, fixed prediction tasks, a budgeted decision replay, reference methods, and… See the full description on the dataset page: https://huggingface.co/datasets/kobzaond/RLVRAMBench.

sourceHugging Facemitupdated 10d agoView on Hugging Face

0likes224downloads

rootRIGHTS.md

RIGHTS.md32 linesDownload Raw Back to root

1# Rights and reuse2 3The repository owner selected **MIT** on 15 September 2026 for the authors'4original RLVRAMBench research data and reproduction software.5The accompanying MIT license covers the authors' original benchmark6measurements, curated tables, analysis/evaluation code, and benchmark7documentation to the extent the authors hold the relevant rights.8 9This is not a blanket relicensing of the evidence archive or upstream10material. In the larger research repository the applicable license text is11`LICENSE-RLVRAMBench`; in the standalone code and dataset repositories it12is `LICENSE`. The original paper remains an authored manuscript;13this notice does not select a journal publication license on the authors'14behalf.15 16Third-party source code, prepared or quoted upstream dataset content,17model descriptions, and the IEEE manuscript template retain their own18terms and notices. In particular, the instrumented VERL runtime and the19TransferQueue dependency retain the license files in their respective20directories. Dataset attribution and recorded upstream terms accompany21the archived evidence. Users must follow those terms for the corresponding22material, including prepared workload examples.23 24No upstream model weights, trained checkpoint shards, or optimizer-state25payloads are redistributed in this release. Their exact identities and26versions are recorded for users to obtain under the upstream terms.27 28The benchmark and reproduction-code repositories are public and ungated.29No DOI is asserted until the authors create one. License selection is30separate from author submission approval, funding/conflict declarations,31and any final journal paperwork.32