RLVRAMBench — Electronics submission

This is the dataset accompanying Memory Feasibility in Colocated Language-Model Reinforcement Learning: A Failure-Aware Measurement Study, prepared for the MDPI Electronics revision on 23 October 2026.

Submission version: electronics-submission-2026-10-23-r2. This version is published on main and pinned by the same tag. Dataset DOI: 10.57967/hf/10780. Earlier releases remain in the repository history and releases/.

• Main manuscript
• Scientific supplement
• Point-by-point reviewer response
• Editable TeX source ZIP
• Matching reproduction code
• Release manifest and archive hashes
• Submission data dictionary
• Reconstruction guide

Authors: Ondřej Kobza and Jan Šedivý, Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague. Author ORCIDs: Kobza, Šedivý.

What the dataset measures

RLVRAMBench measures GPU-memory feasibility when language-model generation and reinforcement-learning updates share devices. It retains successful, memory-failing, and ineligible executions. It provides stage measurements, controlled changes to execution settings, repeated configuration labels, frozen prediction tasks, and raw evidence. It does not measure model intelligence, convergence, or full-training quality.

The main experiments use synchronous LoRA-GRPO with VERL/vLLM on A100-40GB GPUs. Follow-ups add full-parameter 1.5B/3B actors, reservation and state-release sweeps, iteration costs, native LLMem validation, and an MI250X replication. The MI250X panel includes PPO with a full-parameter critic and changes hardware and software together; it is not an isolated hardware comparison.

Submission cohorts

The views above expose separate experimental units. Do not sum their rows as independent experiments.

Cohort	Coverage and interpretation
Historical transfer benchmark	52 tasks, 400 queries reusing 94 distinct targets
Independent A100 offload controls	144 runs; 48 three-seed configurations: 36 within margin, 12 memory failures
A100 adaptation and no-KL controls	108 formal runs: 106 within margin, 2 ineligible
A100 reservation sweep	144 runs: 111 within margin, 11 above margin, 21 memory failures, 1 ineligible
A100 state-release sweep	36 runs: 17 within margin, 6 above margin, 12 memory failures, 1 ineligible
A100 iteration costs	24 runs, all within margin; 20 measured updates per run
Native LLMem validation	24 prediction/measurement pairs in the supported native training setting
MI250X	462 planned slots: 363 completions, 98 memory failures, 1 ineligible outcome; 154 configurations
DeepSpeed comparison	Recorded estimator calls and decisions on the same 400 benchmark queries

The default view contains the 154 MI250X configurations, with panel identifying the five substudies. Its A100 replay has 79 agreements among 89 resolved configurations, plus one unresolved configuration. The unresolved Phi-4-mini/GSM8K c5 setting still exposes its two diagnosed OOMs. The MI250X ledger records 510 started jobs and 511 executions, including one Slurm restart. Pilots and unstarted submissions are separately identified.

Outcome rules and units

Memory is the maximum used memory on any one allocated device, in MiB, not a sum across devices. Completed-run peaks require successful completion and valid measurement evidence. An observed peak from a failed run is not its counterfactual completed memory demand.

Label	Meaning
within_margin	All three eligible repetitions complete below or at the operating limit
above_margin	All three complete; at least one exceeds the operating limit
memory_failure	All three repetitions are eligible; at least one has a diagnosed device-memory failure
unresolved	Fewer than three eligible repetitions; known adverse observations remain visible

The operating limit is capacity minus 2,048 MiB: 38,912 MiB for the A100 devices and 63,472 MiB for MI250X GCDs. This is a declared operating convention, not a probability guarantee. Per-run status vocabularies differ between historical collectors; see the data dictionaries. Missing values are missing evidence or inapplicable quantities, not zero memory.

Load a view

The direct Hub-loading interface is tested with Python 3.11, datasets==5.1.0 and huggingface_hub==2.2.0. Install these from dataset-client-requirements.txt in a separate reader environment. Older clients such as datasets==2.12.0 ignore the card's configuration mapping and may try to combine unrelated audit JSON files. The frozen raw-analysis environment is documented separately in the reproduction guide.

python -m pip install -r dataset-client-requirements.txt

from datasets import load_dataset

configurations = load_dataset(
    "kobzaond/RLVRAMBench",
    "mi250x_configurations",
    revision="electronics-submission-2026-10-23-r2",
    split="measured",
)

Other views retain controls, historical transfer tasks, A100 follow-ups, MI250X raw-derived trials, scheduler accounting, and estimator comparisons. Use the configuration selector above or the inventory in verification/dataset-views.json.

Raw evidence and reproduction

All four required raw-evidence archives are available under releases/:

1. RLVRAMBench-review-artifact-submission-20260917-b0018d0.tar.zst: historical measurements, admission/estimation, and replications.
2. RLVRAMBench-offload-raw-20261022.tar.zst: independent offload controls, frozen source, and allocation accounting.
3. electronics-full-20261021-a53af63.tar.zst: the four A100 follow-up panels, development attempts, native LLMem, and their frozen execution sources.
4. RLVRAMBench-mi250x-raw-20261022.tar.zst: MI250X logs, device maps, memory/phase traces, submissions, and amendments.

Each archive has a SHA-256 sidecar. release.json records archive hashes, coverage, and the code revision. MANIFEST.sha256 covers the submission files. The reproduction guide separates raw reconstruction from replay of already derived tables. No new GPU experiments are needed to reproduce the released analyses. Verification reports describe author-performed checks; they do not claim external replication.

The electronics_revision_2026_10/ folder is retained as an earlier curated snapshot. The root-level paths and version tag above define this submission. Historical IEEE PDFs remain archived; the linked MDPI PDFs are the current submission documents.

Evaluation contract and limitations

Transfer queries expose source evidence and target settings. Target outcomes are public for audit but must not be read when producing predictions for that task. Independent screening and evaluation seeds remain distinct. Report failure approvals, above-margin approvals, useful approvals, and coverage together; aggregate accuracy alone is insufficient.

Measurements depend on the specified stack, device capacity, model, workload, state-release behavior, and requested duration. The longest runs cover 100 steps; the additional MI250X duration panel covers 40. The evidence does not establish frontier-scale feasibility, multi-node scaling, behavior on other engines or accelerators, or full-schedule convergence. The native LLMem and out-of-scope DeepSpeed component comparisons have different purposes. See the paper for scope and attribution limits.

License and citation

Original benchmark code, measurements, and documentation use the MIT license. Third-party material retains its original terms; see RIGHTS.md. Model weights and optimizer checkpoints are excluded. Source datasets and model identifiers are provenance, not a relicensing of upstream assets.

Use CITATION.cff, the dataset DOI, and the version tag when citing this release. The companion manuscript is a journal submission; no acceptance or publication status is implied.