RLVRAMBench / verification /layout-audit.json
kobzaond's picture
Explain the Figure 1 dash and unresolved telemetry in the caption
00e61d8 verified
Raw History Blame Contribute Delete
24.3 kB
{
"release": "electronics-submission-2026-10-23-r2",
"documents": {
"RLVRAMBench_MDPI": {
"sha256": "e561df8e88ae52c6d091afecb4146df10b17825d27866f5db10698c185b4ba00",
"pages": 43,
"figures": 7,
"tables": 21,
"captions": [
{
"kind": "Table",
"number": "1",
"page": 4,
"x": 166.098,
"y": 363.774058,
"first_line": "Table 1. Different questions addressed by related memory analyses. The last column describes the"
},
{
"kind": "Table",
"number": "2",
"page": 6,
"x": 166.098,
"y": 74.421058,
"first_line": "Table 2. Key terms used in this article."
},
{
"kind": "Table",
"number": "3",
"page": 10,
"x": 166.098,
"y": 74.421058,
"first_line": "Table 3. How the experiments support the research questions."
},
{
"kind": "Table",
"number": "4",
"page": 11,
"x": 166.098,
"y": 74.421058,
"first_line": "Table 4. Workload settings for the boundary and batch/reservation experiments."
},
{
"kind": "Table",
"number": "5",
"page": 11,
"x": 166.098,
"y": 335.939058,
"first_line": "Table 5. Configuration levels used in the boundary experiment. Each level bundles three settings."
},
{
"kind": "Table",
"number": "6",
"page": 12,
"x": 166.098,
"y": 74.421058,
"first_line": "Table 6. Recorded attempts for the original 588 planned configuration\u2013seed slots. Completed runs"
},
{
"kind": "Table",
"number": "7",
"page": 17,
"x": 166.098,
"y": 113.076058,
"first_line": "Table 7. Additional RL panels. Counts are planned slots. The external LLMem validation is a separate"
},
{
"kind": "Table",
"number": "8",
"page": 20,
"x": 166.098,
"y": 184.528058,
"first_line": "Table 8. Outcomes of the independent-control experiment on four GPUs. Each cell contains twelve"
},
{
"kind": "Figure",
"number": "1",
"page": 21,
"x": 166.394,
"y": 346.024058,
"first_line": "Figure 1. Reservation sensitivity with actor micro-batch 8 and generation-state release enabled. Cells"
},
{
"kind": "Table",
"number": "9",
"page": 22,
"x": 166.098,
"y": 74.421058,
"first_line": "Table 9. All base configurations in the release experiment. Prompt/response limits are tokens; their"
},
{
"kind": "Figure",
"number": "2",
"page": 23,
"x": 166.394,
"y": 384.413058,
"first_line": "Figure 2. Micro-batch effects in the same-device replication at each reservation. Each point is a seed\u2019s"
},
{
"kind": "Table",
"number": "10",
"page": 23,
"x": 166.098,
"y": 441.678058,
"first_line": "Table 10. Same-device micro-batch/reservation replication with fixed full instrumentation. Entries are"
},
{
"kind": "Figure",
"number": "3",
"page": 24,
"x": 166.394,
"y": 350.004058,
"first_line": "Figure 3. Targeted batch-by-logging control. Each connected pair compares the actor-stage batch-"
},
{
"kind": "Figure",
"number": "4",
"page": 26,
"x": 166.394,
"y": 346.972058,
"first_line": "Figure 4. Full minus LoRA sampled memory peaks in the native-KL panel. Small points are seed"
},
{
"kind": "Table",
"number": "11",
"page": 26,
"x": 166.098,
"y": 490.050058,
"first_line": "Table 11. Alternative minus anchor in the twenty-update iteration-cost panel, three same-device"
},
{
"kind": "Table",
"number": "12",
"page": 27,
"x": 166.098,
"y": 430.803058,
"first_line": "Table 12. Frozen approval decisions for the 48 configurations of the independent-control experiment."
},
{
"kind": "Table",
"number": "13",
"page": 28,
"x": 166.098,
"y": 168.787058,
"first_line": "Table 13. Frozen decisions on the complete three-repeat labels of the reservation-sensitivity panel."
},
{
"kind": "Figure",
"number": "5",
"page": 29,
"x": 166.394,
"y": 372.136058,
"first_line": "Figure 5. Three-state labels across the boundary experiment. Each cell combines three fresh processes."
},
{
"kind": "Table",
"number": "14",
"page": 29,
"x": 166.098,
"y": 416.579058,
"first_line": "Table 14. Copying labels from a source model to a target model at the primary margin. Each direction"
},
{
"kind": "Table",
"number": "15",
"page": 30,
"x": 166.098,
"y": 157.227058,
"first_line": "Table 15. Outcomes at the same levels on two and four GPUs. \u201cWithin\u201d and \u201cabove\u201d are completed"
},
{
"kind": "Table",
"number": "16",
"page": 30,
"x": 166.098,
"y": 602.011058,
"first_line": "Table 16. Admission results for four prospective cases. Entries under the rules give approved /"
},
{
"kind": "Table",
"number": "17",
"page": 32,
"x": 166.098,
"y": 168.253058,
"first_line": "Table 17. LLMem native-training comparison, three pairs per cell. Peaks are in MiB; relative error is"
},
{
"kind": "Table",
"number": "18",
"page": 33,
"x": 166.098,
"y": 74.421058,
"first_line": "Table 18. One-step and 100-step outcomes, with memory measurements from completed long runs."
},
{
"kind": "Table",
"number": "19",
"page": 33,
"x": 166.098,
"y": 559.808058,
"first_line": "Table 19. Repeated labels of the 90 replayed A100 configurations on MI250X. Each label combines"
},
{
"kind": "Table",
"number": "20",
"page": 35,
"x": 166.098,
"y": 74.421058,
"first_line": "Table 20. Additional MI250X panels. Labels combine three fresh processes; \u201cwake-up\u201d denotes"
},
{
"kind": "Figure",
"number": "6",
"page": 36,
"x": 166.394,
"y": 330.827058,
"first_line": "Figure 6. Diagnosed memory failures by experiment. In the plot labels, \u201cState release\u201d denotes the"
},
{
"kind": "Figure",
"number": "7",
"page": 37,
"x": 166.394,
"y": 668.457058,
"first_line": "Figure 7. Agreement of individual within-margin approval decisions across run lengths, by protocol,"
},
{
"kind": "Table",
"number": "21",
"page": 38,
"x": 166.098,
"y": 476.639058,
"first_line": "Table 21. Limits on extrapolating the measured memory boundaries."
}
],
"substantial_word_overlap_candidates": [
{
"page": 32,
"words": [
"M",
"\u02c6\ufe01"
],
"fraction": 0.818
}
],
"caption_left_edge_range_pt": [
166.098,
166.394
],
"unresolved_overlap_candidates": [],
"overlap_review": "The hat accent over M in the estimation-error equation is intentional mathematical typesetting."
},
"RLVRAMBench_supplement": {
"sha256": "35fcae54892f4ff27a272fd30dae269a932ddf3453f0ae230ef5ad1101f07e03",
"pages": 38,
"figures": 3,
"tables": 46,
"captions": [
{
"kind": "Table",
"number": "S1",
"page": 2,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S1: Completion and margin compliance in the original boundary experiment. The last column gives"
},
{
"kind": "Table",
"number": "S2",
"page": 3,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S2: Stage contrasts in external memory (MiB) in the instrumentation control. Each entry is full"
},
{
"kind": "Table",
"number": "S3",
"page": 3,
"x": 62.362,
"y": 564.17452,
"first_line": "Table S3: Factorial effects per workload (MiB), from nine blocks per workload. All runs use full instrumentation."
},
{
"kind": "Figure",
"number": "S1",
"page": 4,
"x": 62.362,
"y": 474.95152,
"first_line": "Figure S1: Supporting summaries of the memory mechanisms. The panel for the release experiment distinguishes"
},
{
"kind": "Table",
"number": "S4",
"page": 5,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S4: Realized sequence lengths in the 100-step duration tests, in tokens. \u201cCap batches\u201d counts batches with"
},
{
"kind": "Table",
"number": "S5",
"page": 5,
"x": 62.362,
"y": 516.66352,
"first_line": "Table S5: Effects in the earlier 100-step duration test, by case (MiB). Brackets give observed ranges across runs,"
},
{
"kind": "Figure",
"number": "S2",
"page": 6,
"x": 62.362,
"y": 299.19852,
"first_line": "Figure S2: First-step and maximum peaks for the twelve completed runs of the earlier 100-step duration test."
},
{
"kind": "Table",
"number": "S6",
"page": 7,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S6: All usable control blocks. Entries are actor-stage effects on the external peak, in MiB. An asterisk marks"
},
{
"kind": "Table",
"number": "S7",
"page": 7,
"x": 62.362,
"y": 261.64752,
"first_line": "Table S7: Realized response lengths for the blocks in Table S6. Entries are mean / maximum tokens, rounded to"
},
{
"kind": "Table",
"number": "S8",
"page": 8,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S8: Descriptive aggregate effects with 95% resampling intervals. Memory effects are in MiB unless stated"
},
{
"kind": "Table",
"number": "S9",
"page": 8,
"x": 62.362,
"y": 493.53052,
"first_line": "Table S9: Calibration configurations. Reservation is the generation reservation, the fraction of GPU memory"
},
{
"kind": "Table",
"number": "S10",
"page": 9,
"x": 62.362,
"y": 263.58652,
"first_line": "Table S10: How missing admission evidence is treated. A recorded invocation is charged even when it does not"
},
{
"kind": "Table",
"number": "S11",
"page": 10,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S11: All prospective candidates. Peaks are in MiB and refer to completed processes; the evaluation peak is"
},
{
"kind": "Table",
"number": "S12",
"page": 11,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S12: Fixed training controls. BF16 is 16-bit bfloat16, and FP32 is 32-bit floating point. Fully Sharded Data"
},
{
"kind": "Table",
"number": "S13",
"page": 11,
"x": 62.362,
"y": 489.89352,
"first_line": "Table S13: Realized work in the boundary experiment. Qwen, Phi, and Granite denote Qwen2.5-3B, Phi-4-mini,"
},
{
"kind": "Table",
"number": "S14",
"page": 12,
"x": 62.362,
"y": 312.67052,
"first_line": "Table S14: Diagnostic rules on the 36 GPU-count transfer queries. \u201cAlways reject\u201d predicts memory failure for"
},
{
"kind": "Table",
"number": "S15",
"page": 13,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S15: Duration settings. Caps are in tokens; every final-step operation listed here is required for completion."
},
{
"kind": "Table",
"number": "S16",
"page": 13,
"x": 62.362,
"y": 306.13952,
"first_line": "Table S16: Retrospective sensitivity to the margin. \u201cCapacity allowed\u201d is the share of device capacity used as the"
},
{
"kind": "Table",
"number": "S17",
"page": 13,
"x": 62.362,
"y": 508.36352,
"first_line": "Table S17: Instrumentation control. Count cells give external-only; full instrumentation, each out of three. Peak"
},
{
"kind": "Table",
"number": "S18",
"page": 16,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S18: Fitting costs and regression peak errors with one model family held out. Each fitting pool has sixty"
},
{
"kind": "Table",
"number": "S19",
"page": 16,
"x": 62.362,
"y": 439.43752,
"first_line": "Table S19: Predictions with one model family held out. Each row evaluates thirty configurations, and the other"
},
{
"kind": "Table",
"number": "S20",
"page": 17,
"x": 62.362,
"y": 344.64852,
"first_line": "Table S20: Original larger-model panel (7B). Each entry gives the number of eligible seeds out of three and the"
},
{
"kind": "Table",
"number": "S21",
"page": 20,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S21: Configuration peaks, in MiB, at four GPUs for levels c1 and c2 in the larger-model replication. Every"
},
{
"kind": "Table",
"number": "S22",
"page": 20,
"x": 62.362,
"y": 400.12152,
"first_line": "Table S22: Predictions by model and actor-offload group (c1 offloads actor parameters; c2/c3 keep them on the"
},
{
"kind": "Table",
"number": "S23",
"page": 21,
"x": 62.362,
"y": 111.52252,
"first_line": "Table S23: Frozen predictions in the larger-model replication. Qwen has eighteen resolved configurations (four"
},
{
"kind": "Table",
"number": "S24",
"page": 22,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S24: Prediction results with independently varied controls, shown separately by model. Each model"
},
{
"kind": "Table",
"number": "S25",
"page": 23,
"x": 62.362,
"y": 417.33452,
"first_line": "Table S25: Observed contrast ranges within blocks, in MiB. Brackets give the minimum and maximum over n"
},
{
"kind": "Table",
"number": "S26",
"page": 25,
"x": 62.362,
"y": 243.07752,
"first_line": "Table S26: Effect of actor offload (enabled minus disabled) at generation reservation 0.45. Each entry is the"
},
{
"kind": "Table",
"number": "S27",
"page": 25,
"x": 62.362,
"y": 388.91952,
"first_line": "Table S27: Whole-run offload contrasts by case at generation reservation 0.45. Each entry is the three-pair mean"
},
{
"kind": "Table",
"number": "S28",
"page": 28,
"x": 62.362,
"y": 63.75752,
"first_line": "Table S28: Verified actor parameter inventories before FSDP wrapping. Stored parameter tensors are FP32"
},
{
"kind": "Table",
"number": "S29",
"page": 28,
"x": 62.362,
"y": 199.88252,
"first_line": "Table S29: Full minus LoRA peak differences, native KL. Values are mean [minimum, maximum] across jointly"
},
{
"kind": "Table",
"number": "S30",
"page": 28,
"x": 62.362,
"y": 451.05052,
"first_line": "Table S30: Actor offload enabled minus disabled, native KL. Each comparison keeps adaptation mode and"
},
{
"kind": "Table",
"number": "S31",
"page": 29,
"x": 62.362,
"y": 63.00752,
"first_line": "Table S31: Full minus LoRA peak differences, no-KL control. Values are mean [minimum, maximum] across jointly"
},
{
"kind": "Table",
"number": "S32",
"page": 29,
"x": 62.362,
"y": 172.75652,
"first_line": "Table S32: Actor offload enabled minus disabled, no-KL control. Each comparison keeps adaptation mode and"
},
{
"kind": "Table",
"number": "S33",
"page": 29,
"x": 62.362,
"y": 403.44852,
"first_line": "Table S33: Reservation sensitivity. Each cell reports within margin/above margin/memory failure/unresolved"
},
{
"kind": "Table",
"number": "S34",
"page": 30,
"x": 62.362,
"y": 63.00752,
"first_line": "Table S34: Frozen prediction decisions by reservation scope. Scored/planned counts retain unresolved targets;"
},
{
"kind": "Table",
"number": "S35",
"page": 30,
"x": 62.362,
"y": 457.26352,
"first_line": "Table S35: State-release sensitivity at the two GSM8K anchors. Outcome counts are within margin/above"
},
{
"kind": "Table",
"number": "S36",
"page": 31,
"x": 62.362,
"y": 63.00752,
"first_line": "Table S36: Iteration costs for Qwen-7B/GSM8K and Mistral-7B/longer-prompt code. Counts are within"
},
{
"kind": "Table",
"number": "S37",
"page": 31,
"x": 62.362,
"y": 249.02052,
"first_line": "Table S37: Observed work and host memory for the same iteration-cost runs. Token rates use prompt plus"
},
{
"kind": "Table",
"number": "S38",
"page": 31,
"x": 62.362,
"y": 435.03352,
"first_line": "Table S38: Same-device timing contrasts against the offload-on, reservation-0.45, release-on anchor. Time"
},
{
"kind": "Figure",
"number": "S3",
"page": 32,
"x": 62.362,
"y": 330.64552,
"first_line": "Figure S3: Mean iteration time of each run over updates 3\u201320. Points are seeds, diamonds are condition means,"
},
{
"kind": "Table",
"number": "S39",
"page": 32,
"x": 62.362,
"y": 501.46352,
"first_line": "Table S39: GPU-hours used by the follow-up experiments. The allocation and task columns overlap and must not"
},
{
"kind": "Table",
"number": "S40",
"page": 33,
"x": 62.362,
"y": 97.57952,
"first_line": "Table S40: MI250X slots by panel. Executions count every started execution, including amendment reruns and one Slurm restart after a node failure; the latest attempt gives"
},
{
"kind": "Table",
"number": "S41",
"page": 33,
"x": 62.362,
"y": 236.73652,
"first_line": "Table S41: Every replayed A100 configuration. Peaks are maxima over completed repetitions (MiB); failure stages count failing processes (A100 / MI250X). Off = actor offload"
},
{
"kind": "Table",
"number": "S41",
"page": 34,
"x": 62.362,
"y": 62.67552,
"first_line": "Table S41: Continued."
},
{
"kind": "Table",
"number": "S41",
"page": 35,
"x": 62.362,
"y": 62.67552,
"first_line": "Table S41: Continued."
},
{
"kind": "Table",
"number": "S42",
"page": 36,
"x": 62.362,
"y": 63.00752,
"first_line": "Table S42: 7B configurations: A100 and MI250X repeated labels."
},
{
"kind": "Table",
"number": "S43",
"page": 36,
"x": 62.362,
"y": 390.28952,
"first_line": "Table S43: Reservation sweep at c3 settings (GSM8K, two GCDs, LoRA-GRPO, one step). Times are per"
},
{
"kind": "Table",
"number": "S44",
"page": 37,
"x": 62.362,
"y": 63.00752,
"first_line": "Table S44: GRPO-LoRA, PPO with a full-parameter critic, and full-parameter GRPO (GSM8K, two GCDs, one"
},
{
"kind": "Table",
"number": "S45",
"page": 37,
"x": 62.362,
"y": 285.56252,
"first_line": "Table S45: One-step screens and 40-step runs of the A100 temporal configurations (LoRA-GRPO). Times are per"
},
{
"kind": "Table",
"number": "S46",
"page": 38,
"x": 62.362,
"y": 377.49752,
"first_line": "Table S46: Estimator decisions on all benchmark targets. Recall is within-margin recall; approval means a"
}
],
"substantial_word_overlap_candidates": [],
"unresolved_overlap_candidates": []
},
"RLVRAMBench_reviewer_response": {
"sha256": "124c501e804fef58220812cabdf3bb61a0073044551c023b35c637b2f23953af",
"pages": 6,
"figures": 0,
"tables": 0,
"captions": [],
"substantial_word_overlap_candidates": [],
"unresolved_overlap_candidates": []
}
},
"new_gpu_experiments": false,
"status": "passed",
"reviewed_at_utc": "2026-10-23T12:59:11.490431+00:00",
"visual_review": {
"all_pages_inspected": true,
"main_figures": 7,
"main_tables": 21,
"supplement_figures": 3,
"supplement_tables": 46,
"main_figure_width": "manuscript text column",
"main_captions_left_aligned": true,
"table_headers_and_numeric_cells": "aligned; no clipped or overlapping cells observed",
"continued_inventory": "MI250X replay uses three landscape pages with repeated headers",
"short_tables_kept_together": [
"S34",
"S44"
],
"all_main_float_labels_referenced": true
},
"material_build_warnings": [],
"source_zip_standalone_rebuild": true,
"data_preservation_report": "layout-data-preservation.json",
"figure1_clarification": "Dash marks the one invalid-telemetry run; caption identifies its configuration and exclusion from the counts.",
"float_page_changes_since_previous_release": [
[
"RLVRAMBench_MDPI",
[
"Table",
"9"
],
21,
22
],
[
"RLVRAMBench_MDPI",
[
"Figure",
"4"
],
25,
26
],
[
"RLVRAMBench_MDPI",
[
"Table",
"18"
],
32,
33
]
]
}