kobzaond commited on
Commit
7bae491
·
verified ·
1 Parent(s): c9bd99f

Release critically revised manuscript and fully audited batch-by-logging control

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +1 -0
  2. CITATION.cff +1 -1
  3. README.md +25 -14
  4. paper/RLVRAMBench_IEEE_Access_source.zip +2 -2
  5. paper/RLVRAMBench_IEEE_Access_submission.pdf +2 -2
  6. paper/RLVRAMBench_IEEE_Access_submission.tex +1029 -0
  7. paper/RLVRAMBench_supplement.pdf +3 -0
  8. paper/RLVRAMBench_supplement.tex +347 -0
  9. paper/figures/review_instrumentation_batch.pdf +0 -0
  10. paper/figures/review_instrumentation_batch.png +3 -0
  11. paper/figures/standard_grpo_boundary.pdf +0 -0
  12. paper/figures/standard_grpo_boundary.png +3 -0
  13. paper/figures/standard_grpo_mechanisms.pdf +0 -0
  14. paper/figures/standard_grpo_mechanisms.png +3 -0
  15. paper/figures/standard_grpo_temporal.pdf +0 -0
  16. paper/figures/standard_grpo_temporal.png +3 -0
  17. release.json +15 -8
  18. releases/RLVRAMBench-release-critical-revision-20260914-328c1c3.json +24 -0
  19. releases/RLVRAMBench-review-artifact-critical-revision-20260914-328c1c3.tar.zst +3 -0
  20. releases/RLVRAMBench-submission-scientific-package-critical-revision-20260914-328c1c3.zip +3 -0
  21. tables/review_revision/attempt-flow.csv +10 -0
  22. tables/review_revision/attempt-workload-summary.json +190 -0
  23. tables/review_revision/control-allocation-attempts.csv +9 -0
  24. tables/review_revision/control-blocks.csv +7 -0
  25. tables/review_revision/control-cases.csv +3 -0
  26. tables/review_revision/control-process-attempts.csv +33 -0
  27. tables/review_revision/control-provenance.csv +7 -0
  28. tables/review_revision/control-realized-work.csv +25 -0
  29. tables/review_revision/control-summary.json +134 -0
  30. tables/review_revision/control-trials.csv +25 -0
  31. tables/review_revision/corpus-weighting-sensitivity.csv +10 -0
  32. tables/review_revision/decision-records.csv +397 -0
  33. tables/review_revision/decision-summaries.csv +21 -0
  34. tables/review_revision/factorial-allocation-history.csv +145 -0
  35. tables/review_revision/factorial-peak-stages.csv +145 -0
  36. tables/review_revision/factorial-workload-effects.csv +5 -0
  37. tables/review_revision/instrumentation-stage-cells.csv +13 -0
  38. tables/review_revision/instrumentation-stage-pairs.csv +37 -0
  39. tables/review_revision/instrumentation-stage-trials.csv +73 -0
  40. tables/review_revision/long-run-exposure.csv +7 -0
  41. tables/review_revision/original-temporal-case-effects.csv +3 -0
  42. tables/review_revision/realized-workload-cases.csv +53 -0
  43. tables/review_revision/realized-workload-trials.csv +0 -0
  44. tables/review_revision/recorded-attempts.csv +0 -0
  45. tables/review_revision/release-idle-baselines.csv +37 -0
  46. tables/review_revision/state-release-design.csv +7 -0
  47. tables/review_revision/summary.json +617 -0
  48. upload-manifest.json +185 -15
  49. verification/VALIDATION.md +6 -1
  50. verification/revised-reproduction.json +16 -0
.gitattributes CHANGED
@@ -59,3 +59,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
  paper/RLVRAMBench_IEEE_Access_submission.pdf filter=lfs diff=lfs merge=lfs -text
 
 
59
  *.mp4 filter=lfs diff=lfs merge=lfs -text
60
  *.webm filter=lfs diff=lfs merge=lfs -text
61
  paper/RLVRAMBench_IEEE_Access_submission.pdf filter=lfs diff=lfs merge=lfs -text
62
+ paper/RLVRAMBench_supplement.pdf filter=lfs diff=lfs merge=lfs -text
CITATION.cff CHANGED
@@ -2,7 +2,7 @@ cff-version: 1.2.0
2
  message: "Please cite the RLVRAMBench evidence release and associated paper."
3
  type: dataset
4
  title: "RLVRAMBench: benchmark evidence release"
5
- version: "2026.09.14"
6
  date-released: "2026-09-14"
7
  authors:
8
  - family-names: Kobza
 
2
  message: "Please cite the RLVRAMBench evidence release and associated paper."
3
  type: dataset
4
  title: "RLVRAMBench: benchmark evidence release"
5
+ version: "2026.09.14-r2"
6
  date-released: "2026-09-14"
7
  authors:
8
  - family-names: Kobza
README.md CHANGED
@@ -20,8 +20,8 @@ configs:
20
 
21
  # RLVRAMBench
22
 
23
- Benchmark evidence accompanying **RLVRAMBench: A Failure-Inclusive Benchmark
24
- for GPU-Memory Feasibility in Colocated LoRA GRPO**.
25
 
26
  Authors: Ondřej Kobza and Jan Šedivý, Czech Institute of Informatics, Robotics
27
  and Cybernetics, Czech Technical University in Prague.
@@ -39,15 +39,20 @@ and file-integrity manifests.
39
  - 588 principal processes: 438 historical and 150 prospective revision runs.
40
  - 16 historical source-screen processes, reported separately.
41
  - 40 published CSV/JSON result files in `tables/`.
42
- - The frozen evidence archive under `releases/`: 19,226 manifest entries.
43
- - The 19-page IEEE Access manuscript snapshot accompanying the initial evidence
44
- deposit and its matching LaTeX source bundle under `paper/`.
45
-
46
- The archive is rebuilt from source commit
47
- `289747694f1a71052acc7977e742bfc0a45286a2`. It retains historical material for
48
- provenance. The files under `paper/` accompany that evidence snapshot; the
49
- latest manuscript, including the public repository links, is maintained in
50
- the original [paper repository](https://github.com/kobzaond/rl/tree/rlvrambench-paper/paper/submission).
 
 
 
 
 
51
 
52
  No model weights, trained checkpoint shards, or optimizer-state payloads
53
  are included. Those are distinct model-release artifacts.
@@ -59,8 +64,8 @@ Standalone CPU reproduction code:
59
 
60
  The code repository pins the verified dataset commit, archive filename, and
61
  SHA-256 digest. It verifies extraction, rebuilds the complete result set from
62
- raw evidence without precomputed principal tables, compares all 40 result
63
- files, and regenerates four figures. GPU training is not rerun.
64
 
65
  `release.json` identifies the source snapshot and archive digest.
66
  `upload-manifest.json` identifies the size and SHA-256 of each uploaded file,
@@ -75,12 +80,18 @@ general equivalence, or portability beyond the tested settings. Failures,
75
  infrastructure exclusions, instrumentation controls, phase-local effects,
76
  and bounded temporal validation are retained in the evidence.
77
 
 
 
 
 
 
 
78
  ## Access, licenses, and citation
79
 
80
  This benchmark and the reproduction-code repository are **public** as of
81
  2026-09-14, at the repository owner's request. Downloads require neither
82
  authentication nor access approval. The initial deposit was private; the
83
- frozen evidence archive and its SHA-256 digest are unchanged by public release.
84
 
85
  The authors have not yet selected licenses for their original code or
86
  research outputs. Applicable upstream dataset, software, and template notices
 
20
 
21
  # RLVRAMBench
22
 
23
+ Benchmark evidence accompanying **RLVRAMBench: A Benchmark for Memory
24
+ Feasibility in Colocated Language-Model Reinforcement Learning**.
25
 
26
  Authors: Ondřej Kobza and Jan Šedivý, Czech Institute of Informatics, Robotics
27
  and Cybernetics, Czech Technical University in Prague.
 
39
  - 588 principal processes: 438 historical and 150 prospective revision runs.
40
  - 16 historical source-screen processes, reported separately.
41
  - 40 published CSV/JSON result files in `tables/`.
42
+ - 27 revised audit/control CSV/JSON files in `tables/review_revision/`.
43
+ - The new control: 26 started processes, 24 within-margin completions, and
44
+ six usable four-condition blocks. Two initial noncompletions and the
45
+ subsequent additional allocations are explicitly distinguished.
46
+ - Versioned evidence archives under `releases/`; current file counts and
47
+ checksums are recorded in `release.json`.
48
+ - The final 18-page IEEE Access manuscript, five-page supplement, matching
49
+ LaTeX source ZIP, and four figure assets under `paper/`.
50
+
51
+ The current source commit is pinned in `release.json`. Earlier immutable
52
+ archives remain available, including the initial evidence snapshot from
53
+ commit `289747694f1a71052acc7977e742bfc0a45286a2`. Files under `paper/`
54
+ are synchronized with the direct-TeX manuscript in the
55
+ [paper repository](https://github.com/kobzaond/rl/tree/rlvrambench-paper/paper/submission/ieee_access_submission).
56
 
57
  No model weights, trained checkpoint shards, or optimizer-state payloads
58
  are included. Those are distinct model-release artifacts.
 
64
 
65
  The code repository pins the verified dataset commit, archive filename, and
66
  SHA-256 digest. It verifies extraction, rebuilds the complete result set from
67
+ raw evidence without precomputed principal tables, compares all 40 original
68
+ and 27 revised result files, and regenerates four figures. GPU training is not rerun.
69
 
70
  `release.json` identifies the source snapshot and archive digest.
71
  `upload-manifest.json` identifies the size and SHA-256 of each uploaded file,
 
80
  infrastructure exclusions, instrumentation controls, phase-local effects,
81
  and bounded temporal validation are retained in the evidence.
82
 
83
+ The new control supports the direction of the actor-stage batch contrast
84
+ without allocator logging. Generated responses differ across conditions;
85
+ the contrasts do not isolate tracing overhead on identical tensors or
86
+ establish logging equivalence. All original 657 recorded attempts, including
87
+ 69 exclusions from the 588 retained outcomes, are accounted for.
88
+
89
  ## Access, licenses, and citation
90
 
91
  This benchmark and the reproduction-code repository are **public** as of
92
  2026-09-14, at the repository owner's request. Downloads require neither
93
  authentication nor access approval. The initial deposit was private; the
94
+ earlier frozen evidence archive is preserved alongside the revised release.
95
 
96
  The authors have not yet selected licenses for their original code or
97
  research outputs. Applicable upstream dataset, software, and template notices
paper/RLVRAMBench_IEEE_Access_source.zip CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:2d28eb52f6114ba6f8e21e85cac2add3bd1ee06a6a6b1940cf2076560d70b989
3
- size 1149303
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cb76104ff25aeeded84335cb8f08d9d6fa13d21c203b4f4a100972800a3ab385
3
+ size 1147484
paper/RLVRAMBench_IEEE_Access_submission.pdf CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:8ea9b321c1e94a742d7ce3304f0f455d4a65c7d402ce69a3ec7db2acc9e648a0
3
- size 466017
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:0eda7af5034a3f1fd9f2a531bd13af3ea5854aa9560ae64fb40d80d5914b8bc0
3
+ size 434230
paper/RLVRAMBench_IEEE_Access_submission.tex ADDED
@@ -0,0 +1,1029 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ % Authoritative submission manuscript. Edit this TeX directly; do not overwrite
2
+ % it by regenerating from the earlier Markdown preparation sources.
3
+ \documentclass{ieeeaccess}
4
+ \usepackage{cite}
5
+ \usepackage{amsmath,amssymb,amsfonts}
6
+ \usepackage{graphicx,textcomp,booktabs,array,calc}
7
+ \usepackage[T1]{fontenc}
8
+ \usepackage{iftex}
9
+ \ifPDFTeX
10
+ \DeclareUnicodeCharacter{2192}{\ensuremath{\rightarrow}}
11
+ \DeclareUnicodeCharacter{2212}{\ensuremath{-}}
12
+ \fi
13
+ \usepackage{xurl}
14
+ \usepackage[hidelinks]{hyperref}
15
+ \usepackage{fvextra}
16
+ \usepackage{etoolbox}
17
+ \usepackage[expansion=false]{microtype}
18
+ \DefineVerbatimEnvironment{verbatim}{Verbatim}{breaklines=true,breakanywhere=true,fontsize=\footnotesize}
19
+ \providecommand{\tightlist}{\setlength{\itemsep}{0pt}\setlength{\parskip}{0pt}}
20
+ \newcommand{\passthrough}[1]{#1}
21
+ \makeatletter
22
+ \newsavebox\pandoc@box
23
+ \newcommand*\pandocbounded[1]{\sbox\pandoc@box{#1}%
24
+ \ifdim\wd\pandoc@box>\linewidth\resizebox{\linewidth}{!}{\usebox\pandoc@box}%
25
+ \else\usebox\pandoc@box\fi}
26
+ % Publication volume/year/DOI are assigned by the publisher.
27
+ % Suppress the unassigned fields without changing vendor class files.
28
+ \patchcmd{\@maketitle}
29
+ {{\vss\doifont Digital Object Identifier\space\@doi\vss}\vspace*{7.4mm}\par}
30
+ {}{}{\PackageError{rlvrambench}{Cannot suppress the publisher DOI field}{}}
31
+ \def\rlvramfooter{\hbox to\textwidth{%
32
+ {\footervolfont RLVRAMBench}\hfill%
33
+ {\footerpagefont\thepage}}}
34
+ \let\rlvramoriginalheadings\ps@headings
35
+ \def\ps@headings{\rlvramoriginalheadings
36
+ \let\@oddfoot\rlvramfooter\let\@evenfoot\rlvramfooter}
37
+ \let\rlvramoriginaltitlepage\ps@titlepage
38
+ \def\ps@titlepage{\rlvramoriginaltitlepage
39
+ \let\@oddfoot\rlvramfooter\let\@evenfoot\rlvramfooter}
40
+ \makeatother
41
+ \pagestyle{headings}
42
+ \ifXeTeX
43
+ \special{pdf:mapfile t1-times.map}
44
+ \special{pdf:mapfile t1-formata.map}
45
+ \special{pdf:mapfile t1-giovannistd.map}
46
+ \fi
47
+ \setlength{\emergencystretch}{1em}
48
+ \begin{document}
49
+ \history{}
50
+ \doi{}
51
+ \hypersetup{pdftitle={RLVRAMBench: A Benchmark for Memory Feasibility in Colocated Language-Model Reinforcement Learning},pdfauthor={Ond\v{r}ej Kobza, Jan \v{S}ediv\'{y}}}
52
+ \title{RLVRAMBench: A Benchmark for Memory\\
53
+ Feasibility in Colocated Language-Model\\
54
+ Reinforcement Learning}
55
+ \author{\uppercase{Ond\v{r}ej Kobza}\authorrefmark{1}, AND \uppercase{Jan \v{S}ediv\'{y}}\authorrefmark{1}}
56
+ \address[1]{Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague, Jugosl\'{a}vsk\'{y}ch partyz\'{a}n\r{u} 1580/3, 160 00 Prague 6, Czech Republic}
57
+ \tfootnote{ORCID: Ond\v{r}ej Kobza: \href{https://orcid.org/0000-0002-0529-9860}{0000-0002-0529-9860}; Jan \v{S}ediv\'{y}: \href{https://orcid.org/0000-0003-0626-2303}{0000-0003-0626-2303}}
58
+ \markboth{Kobza and \v{S}ediv\'{y}: RLVRAMBench}{Kobza and \v{S}ediv\'{y}: RLVRAMBench}
59
+ \begin{abstract}
60
+ Reinforcement learning for large language models can share graphics processors
61
+ between response generation and training. Memory demand changes between these
62
+ stages, so successful generation does not establish that training will finish.
63
+ We introduce \textbf{RLVRAMBench}, a benchmark for measuring memory feasibility
64
+ while retaining unsuccessful attempts and their diagnostic evidence. Its
65
+ protocol separates completion from completion with a specified memory margin,
66
+ combines fresh-process repetitions, and tests short screens against longer
67
+ runs. We study colocated low-rank adaptation with Group Relative Policy
68
+ Optimization in a fixed execution stack. Simple model-specific lookup rules
69
+ recover the observed workload labels on the tested grid. Transfer between
70
+ models and hardware allocations is less uniform: some errors concern the
71
+ chosen margin, while others admit configurations that fail. Controlled
72
+ comparisons quantify the effect of retained generation state and show that
73
+ processing more training examples at once can increase training-stage memory
74
+ without increasing the overall peak. Additional controls confirm this direction
75
+ with framework memory logging disabled, while effect sizes vary between
76
+ logging conditions and repetitions. Longer runs reveal margin crossings that
77
+ individual short screens miss. Public records and analysis software support
78
+ reconstruction, including excluded and unresolved attempts. The contribution
79
+ is a reproducible evaluation protocol and a documented set of measurements
80
+ for comparing configuration decisions in colocated training.
81
+ \end{abstract}
82
+ \begin{keywords}
83
+ colocated execution, failure analysis, GPU memory, Group Relative Policy Optimization, performance benchmarking, reproducibility
84
+ \end{keywords}
85
+ \titlepgskip=-21pt
86
+ \maketitle
87
+ \section{Introduction}\label{introduction}
88
+
89
+ Reinforcement learning for large language models alternates between
90
+ generating responses and updating the model that generated them. The
91
+ generated responses are called \emph{rollouts}, and the model being trained
92
+ is the \emph{policy}, also called the actor. A training iteration may
93
+ additionally score responses with a fixed reference policy. Weight
94
+ synchronization copies updated weights to the generation engine. Colocated
95
+ execution assigns these stages to the same graphics processing units
96
+ (GPUs), allowing training and generation to share hardware.
97
+
98
+ Sharing hardware does not make memory demand constant. The generation engine reserves space for model state and intermediate results; some of that state may remain in memory during training or weight transfer. A configuration that completes generation may therefore run out of memory during a later stage. Users need to know not only how much memory a successful run consumed, but also which configurations failed and where. An out-of-memory (OOM) failure is part of the answer, not a missing measurement to discard.
99
+
100
+ This paper studies \emph{memory feasibility}: whether a configuration completes its requested work. We also define a more conservative notion of \emph{operational safety}, which requires completion while leaving a specified memory margin. Here, safety refers only to resource use; it says nothing about output correctness or training convergence. A single successful run does not establish either repeatability or safety over a longer training schedule.
101
+
102
+ RLVRAMBench evaluates these questions for Group Relative Policy Optimization (GRPO) with low-rank adaptation (LoRA). GRPO learns from the relative rewards of responses to the same input prompt, while LoRA limits training to small adapter matrices rather than updating all model weights. The implementation uses VERL, a framework for language-model reinforcement learning, with vLLM as its response-generation engine. We hold this execution design fixed and study the configuration choices available to its users.
103
+
104
+ RLVRAMBench complements systems that improve placement, scheduling, and memory sharing by evaluating what their exposed configuration choices can support. It links completion and margin-compliance decisions to repeated processes, individual execution stages, and the full attempt history.
105
+
106
+ The paper makes four contributions:
107
+
108
+ \begin{enumerate}
109
+ \def\labelenumi{\arabic{enumi}.}
110
+ \tightlist
111
+ \item
112
+ \textbf{A failure-inclusive measurement protocol.} It defines valid trials, distinguishes completion from memory safety, and specifies how repeated runs form a configuration-level result.
113
+ \item
114
+ \textbf{Transparent transfer baselines.} A common six-level configuration grid is repeated across three model families and four workloads. Simple lookup rules expose both the grid's workload invariance and the different consequences of transferring labels between models or GPU counts.
115
+ \item
116
+ \textbf{Stage-level controls and longer-run tests.} Experiments vary retained generation state, the number of training examples processed at once, and generation-memory reservation. Paired controls examine instrumentation, while longer runs include validation and checkpoint writing.
117
+ \item
118
+ \textbf{A public reproduction package.} Raw records, experiment definitions, exclusion decisions, analysis software, and file-integrity checks allow readers to reconstruct the reported results, including failures.
119
+ \end{enumerate}
120
+
121
+ \subsection{Research questions}\label{research-questions}
122
+
123
+ The study addresses three research questions (RQs):
124
+
125
+ \begin{itemize}
126
+ \tightlist
127
+ \item
128
+ \textbf{RQ1 --- Feasibility and transfer:} How do completion and margin-compliance labels transfer between the tested workloads, models, and GPU counts?
129
+ \item
130
+ \textbf{RQ2 --- Memory mechanisms:} How do retained generation state, training batch size, and generation-memory reservation affect completion and stage-specific memory use?
131
+ \item
132
+ \textbf{RQ3 --- Repeatability over time:} Do safety labels agree across fresh processes and remain valid during longer training, validation, and checkpoint writing?
133
+ \end{itemize}
134
+
135
+ \section{Related work}\label{related-work}
136
+
137
+ \subsection{Execution and resource sharing in language-model training}\label{execution-and-resource-sharing-in-language-model-training}
138
+
139
+ Systems for reinforcement learning from human feedback (RLHF) and related post-training methods must coordinate several models and execution stages. DeepSpeed-Chat combines training and inference optimizations \cite{yao2023deepspeedchat}. HybridFlow describes these stages as a dataflow and redistributes actor state between training and generation \cite{sheng2024hybridflow}. RLHFuse combines work within and across stages to improve utilization \cite{zhong2024rlhfuse}. OpenRLHF supports several placements through distributed scheduling \cite{hu2024openrlhf}, while ReaL searches plans that reallocate model parameters during execution \cite{mei2024realhf}.
140
+
141
+ A second group changes the execution relationship between generation and training. AReaL permits asynchronous progress \cite{fu2025areal}, and RollArt separates rollout execution for multi-task, multi-turn training \cite{tan2026rollart}. RobustRL addresses recovery after machine faults rather than memory limits within a working allocation \cite{chen2026robustrl}.
142
+
143
+ Outside reinforcement learning, SIRIUS shares memory between response-time-sensitive inference and background training, adjusting the training batch to protect inference service \cite{wang2025sirius}. Such resource-sharing systems motivate measurements that distinguish an individual stage's demand from the overall peak.
144
+
145
+ RLVRAMBench measures completion and memory-margin compliance within a fixed colocated system. Repeated configuration labels and their transfer are central measurements, rather than recovery from machine faults or improved scheduling.
146
+
147
+ \subsection{Memory management during generation and training}\label{memory-management-during-generation-and-training}
148
+
149
+ Models process text as \emph{tokens}, the units produced by their tokenizer. During generation, a key--value (KV) cache stores intermediate attention results for previously generated tokens. vLLM's PagedAttention manages this cache in blocks to reduce fragmentation, or unusable gaps in allocated memory \cite{kwon2023pagedattention}. Its Sleep Mode can move model state out of GPU memory or discard cached state between generation periods \cite{vllm2026sleepmode}. Hydra-RLHF reduces the memory required by Proximal Policy Optimization (PPO), another policy-training algorithm, through shared base models and selectable LoRA adapters \cite{santacroce2023hydra}.
150
+
151
+ Zhou et al.\ study memory use in DeepSpeed-Chat and ColossalChat, including state retained across inference and training \cite{zhou2024rlhfmemory}. They evaluate releasing unused cache from the framework's memory manager, or \emph{allocator}, between stages and show that a training-memory control can leave the overall peak unchanged when inference determines that peak. We extend this measurement question to colocated LoRA GRPO with repeated configuration labels, retained memory failures, and transfer and execution-horizon tests. The vLLM state offload and discard tested here is distinct from releasing unused PyTorch allocator cache.
152
+
153
+ \subsection{Failure-aware configuration measurement}\label{failure-aware-configuration-measurement}
154
+
155
+ SLO-Guard includes crashes when tuning standalone vLLM serving under service-level objectives (SLOs), such as response-time targets \cite{lysenstoen2026sloguard}. RLVRAMBench applies failure-inclusive measurement to synchronous GRPO, where a run can fail before its full memory peak is observed. We retain the failure location, combine repetitions before assigning a configuration label, and test transfer between complete model--workload settings. These records can support later configuration search, but the present study does not evaluate a search algorithm.
156
+
157
+ \subsection{Closest-scope comparison}\label{closest-scope-comparison}
158
+
159
+ \begin{table*}[!t]
160
+ \centering
161
+ \caption{Scope of the closest systems and measurement work.}
162
+ \footnotesize
163
+ \setlength{\tabcolsep}{3pt}
164
+ \renewcommand{\arraystretch}{1.12}
165
+ \begin{tabular}{@{}
166
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}
167
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}
168
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}
169
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}@{}}
170
+ \toprule\noalign{}
171
+ \begin{minipage}[b]{\linewidth}\raggedright
172
+ Work
173
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
174
+ Evaluated setting
175
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
176
+ Primary intervention or objective
177
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
178
+ Distinction from this benchmark
179
+ \end{minipage} \\
180
+ \midrule\noalign{}
181
+ HybridFlow / RLHFuse \cite{sheng2024hybridflow, zhong2024rlhfuse} & Distributed RLHF & Redistributing state and combining stages & Changes execution to improve efficiency \\
182
+ SIRIUS \cite{wang2025sirius} & Independent inference and training & Memory handover under service constraints & Shares memory between separate workloads \\
183
+ RLHF memory study \cite{zhou2024rlhfmemory} & DeepSpeed-Chat and ColossalChat & State retention, cache release, and stage-dependent peaks & Establishes mechanisms tested here with repeated, failure-inclusive labels \\
184
+ SLO-Guard \cite{lysenstoen2026sloguard} & Standalone language-model serving & Configuration tuning that includes crashes & Selects settings under service constraints \\
185
+ RobustRL \cite{chen2026robustrl} & Post-training machine failures & Recovery of training and generation services & Studies machine faults rather than memory limits \\
186
+ This work & Fixed colocated LoRA GRPO & Repeated, failure-inclusive measurement & Tests configuration safety and its transfer \\
187
+ \bottomrule
188
+ \end{tabular}
189
+ \end{table*}
190
+
191
+ Table 1 summarizes the main evaluated problem in each work. The distinction is the setting and measurement design, not a claim that earlier systems never reported memory failures.
192
+
193
+ \subsection{Repeatability and few-run uncertainty}\label{repeatability-and-few-run-uncertainty}
194
+
195
+ Deep reinforcement-learning evaluations can be sensitive to implementation details, randomness, and small run counts \cite{henderson2018matters}. Uncertainty estimates are therefore important even when average results appear consistent \cite{agarwal2021precipice}. We apply this principle to memory measurements: configurations are specified in advance, each repetition starts in a fresh process, and comparisons preserve the grouping of runs within workloads.
196
+
197
+ \section{RLVRAMBench}\label{rlvrambench}
198
+
199
+ The experiments study synchronous LoRA GRPO on NVIDIA A100 GPUs with 40 GB of memory. Generation and training share the same devices rather than running as separate services. Most experiments use two GPUs; a follow-up study uses four to test whether the observed labels transfer between allocation sizes.
200
+
201
+ The outcomes concern memory use and completion, not reward or model accuracy. The measured boundaries apply to the selected models, workloads, and software versions. Full-parameter training, other learning algorithms, other accelerators, and alternative execution designs require separate evaluation.
202
+
203
+ \subsection{Cases and configurations}\label{cases-and-configurations}
204
+
205
+ A \emph{workload} specifies the task data and the amount of work requested: prompt and response length limits, responses generated per prompt, and training batch size. A \emph{case} combines that workload with a model and GPU count. Within a case, we vary memory and batching controls to obtain different \emph{configurations}. This distinction allows us to compare settings without treating minor variations of the same workload as independent evidence.
206
+
207
+ \subsection{Configuration factors}\label{configuration-factors}
208
+
209
+ The main controls are batch size, generation-memory reservation, and the location of model state. A \emph{micro-batch} is the portion processed at once on one GPU, in contrast to the full training batch. Separate micro-batches control policy updates and the calculation of response log probabilities used in training.
210
+
211
+ The vLLM reservation specifies the fraction of GPU memory available to the generation engine. Length limits and the maximum number of active sequences also constrain its memory demand. \emph{Offloading} moves parameter or optimizer state to host memory. The release setting instead determines whether generation state is removed from GPU memory between generation periods or kept resident. The experiments vary these controls within the workload definitions given below.
212
+
213
+ \subsection{Measurements}\label{measurements}
214
+
215
+ Each trial records the configuration, random seed, software environment, elapsed time, exit status, and any failure. Experiment identifiers link these records to jobs submitted through Slurm, the cluster's job scheduler.
216
+
217
+ External sampling uses the NVIDIA Management Library (NVML), which exposes device-level memory use and utilization, at 100 ms intervals. Host-side execution markers identify initialization, generation, reference scoring, policy update, weight synchronization, validation, and checkpoint writing. Validation evaluates the current policy; checkpoint writing saves training state for later use. Samples receive the current stage marker. These markers do not synchronize the GPUs, so stage attribution follows the observed execution interval rather than the lifetime of each allocation. Unmarked intervals remain identified as idle or unknown.
218
+
219
+ Framework allocator traces record how PyTorch requests and manages memory. They complement external measurements: allocated tensor memory is not the same quantity as total device memory use. A run's external peak is the maximum over sampled times and allocated GPUs; a stage peak restricts that maximum to the corresponding marked interval. We report memory in mebibytes (MiB, \(2^{20}\) bytes) and gibibytes (GiB, \(2^{30}\) bytes).
220
+
221
+ \subsection{Failure taxonomy}\label{failure-taxonomy}
222
+
223
+ We distinguish failures of the experiment setup from failures caused by the tested configuration. Setup failures include missing dependencies, incorrect GPU visibility, and broken instrumentation; they are documented and excluded. Memory failures during initialization, generation, policy update, or weight synchronization remain experimental outcomes. A failure is assigned to a stage when the execution log and trace support that location. An allocator error alone does not establish fragmentation as its cause.
224
+
225
+ \subsection{Workload strata}\label{workload-strata}
226
+
227
+ GSM8K, the Grade School Math 8K dataset \cite{cobbe2021gsm8k}, contains mathematical word problems. MATH contains competition mathematics problems \cite{hendrycks2021math}, and CodeContests contains programming problems \cite{li2022alphacode}. Prompt limits, generated-response limits, and responses per prompt define distinct memory workloads; they are not interchangeable measures of sequence length.
228
+
229
+ For CodeContests, preparation pools its original validation and test records, excludes problems requiring file input/output or lacking test examples, and orders the remaining prompts by their tokenized length using Qwen2.5-1.5B-Instruct \cite{qwen2024qwen2515b}. The shortest and longest 128 eligible prompts form disjoint \emph{standard code} and \emph{longer-prompt code} subsets. Alternating records in each ordered subset form 64 training and 64 validation examples. The longer subset is selected by length, not by fitting a tail distribution.
230
+
231
+ Preparation profiles use Qwen tokenizers and therefore do not describe the exact batches seen by every model. Actual runs apply the selected model's tokenizer and remove prompts above the workload limit. Section IV-C reports these filtering counts and the realized lengths in completed boundary runs. The model comparison consequently holds the workload definition fixed, not the exact token sequences or, after filtering, necessarily the same examples.
232
+
233
+ For code, the reward checks Python syntax and the presence of input/output operations without executing the generated program. It does not use the CodeContests correctness tests. These workloads therefore exercise code-generation memory demand, not a validated method for improving program correctness.
234
+
235
+ \subsection{Artifact validity and retry policy}\label{artifact-validity-and-retry-policy}
236
+
237
+ Each planned \emph{slot} specifies a configuration and a seed. Its eligible outcome requires a distinct trial record, environment record, execution log, stage trace, and per-GPU measurements. Completion requires a successful exit and evidence that the requested final training step was reached. A nonzero exit enters the measured corpus when its evidence supports a memory failure. Other attempts remain in the attempt inventory with their observed outcome and exclusion reason; a missing trial record does not establish either success or OOM.
238
+
239
+ Allocator logging begins when the instrumented worker starts. Early engine failures can therefore occur before any allocator event is recorded. We assess them using the execution log, external measurements, stage trace, and error stack. Allocator traces are also absent by design in the control that disables allocator instrumentation.
240
+
241
+ A setup repair repeats the original slot without changing its scientific settings. Selection accepts its sole eligible outcome after the documented exclusions and rejects histories with multiple unresolved eligible outcomes. The three intentional seed repetitions are distinct slots, not retries.
242
+
243
+ Section IV-C reconciles the original measured corpus with all recorded attempts belonging to its slots. One excluded attempt has a scheduler-confirmed OOM but lacks a usable trial record because its launcher was changed during execution. Its identical replacement also failed; restoring the known earlier failure therefore does not change that configuration's label. This exclusion is reported as an incomplete memory-failure record, not as a successful run or an unexplained setup failure.
244
+
245
+ The September 13 follow-up studies and September 14 targeted control use immutable source copies. Their runners allow a 60-second shutdown period after a diagnosed OOM and impose per-process time limits of twenty minutes for one step and two hours for a long run. The two external-only timeouts in the original instrumentation control are noncompletions with unresolved cause. Section V-G reports their attempt-level sensitivity separately from the eligible-pair comparison. Detailed repair histories accompany the evidence.
246
+
247
+ \subsection{Safety labels and units of analysis}\label{safety-labels-and-units-of-analysis}
248
+
249
+ We combine the three planned seed slots into a configuration-level record only when all have eligible outcomes. Completion requires all three to finish. Peak memory is the maximum over repetitions, and runtime is the mean over successful repetitions. An unresolved slot prevents a complete repeated label; it is not silently omitted.
250
+
251
+ Operational safety additionally requires a per-GPU peak no greater than 38,912 MiB: 95\% of the reported 40,960 MiB capacity, leaving a 2,048 MiB margin. We chose this margin before the experiments as an operating convention, not as an estimate of the physical failure point. A run can therefore complete yet be classified as unsafe. We later examine how the labels change under other margins.
252
+
253
+ Transfer compares the label for one model or GPU allocation with the observed label for another at the same workload and configuration level. A \emph{false-safe} decision accepts a target configuration that is unsafe; a \emph{false-unsafe} decision rejects one that is safe. We report each direction separately and state its denominator. The highest observed safe level summarizes a boundary, without assuming that untested levels are safe.
254
+
255
+ Processes, configurations, and cases serve different purposes. Individual processes support retry decisions and repeatability checks; configurations support safety labels; complete cases support comparisons across workloads. For a controlled intervention, a \emph{base configuration} is the shared setting before the experimental factor is changed. Its repeated and paired measurements remain together in the statistical analysis.
256
+
257
+ The experiment deliberately concentrates on settings near the margin or a failure boundary. Its process-level fractions describe this evidence set, not a deployment population. Section IV-I explains the grouping and finite-case uncertainty calculations.
258
+
259
+ \subsection{Experiment planning and selection}\label{experiment-planning-and-selection}
260
+
261
+ Pilot runs were used to develop the configuration ranges and instrumentation. Each subsequent experiment was specified in an \emph{experiment matrix}: a fixed list of settings and seeds prepared before those runs began. Pilot results are not included in the repeated estimates. The matrices and dated amendments distinguish planning, setup repairs, and later analysis; they are development records, not independent preregistrations.
262
+
263
+ Exploratory work also tested a custom optimization objective. On 12 September 2026, after those outcomes were available, the study was restricted to standard GRPO. This was a post-outcome scope decision, not a planned comparison between algorithms. The excluded observations remain in the evidence package but contribute to none of this paper's counts or conclusions.
264
+
265
+ The follow-up study was specified on 13 September 2026 and added 150 processes for instrumentation controls, longer-run validation, and four-GPU transfer. Its settings were fixed before execution. In contrast, the stage-specific analysis of the batch/reservation experiment and the margin-sensitivity analysis reuse earlier observations and are explicitly post hoc.
266
+
267
+ A critical review prompted an attempt-level audit, simple transfer baselines, and analyses of realized sequence lengths and stage-specific measurement effects. These reuse existing outcomes and are retrospective. A further 24-process control was specified on 14 September before its execution. It tests the training-batch contrast with and without allocator logging in two selected cases; it does not replace the original results.
268
+
269
+ Configurations within one model--workload case are closely related. Any future prediction study using these data should keep whole cases together when separating training and evaluation data; otherwise, nearly identical settings could appear in both. The present cross-model experiments measure transfer directly and do not train a predictor.
270
+
271
+ \subsection{Outcome summary}\label{outcome-summary}
272
+
273
+ Table 2 summarizes the outcomes and how each treats failed trials.
274
+
275
+ \begin{table*}[!t]
276
+ \centering
277
+ \caption{Endpoints in the eligible corpus. Other and unresolved attempts remain in the separate attempt inventory.}
278
+ \footnotesize
279
+ \setlength{\tabcolsep}{3pt}
280
+ \renewcommand{\arraystretch}{1.12}
281
+ \begin{tabular}{@{}
282
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}
283
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}
284
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}
285
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2500}}@{}}
286
+ \toprule\noalign{}
287
+ \begin{minipage}[b]{\linewidth}\raggedright
288
+ Endpoint
289
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
290
+ Scientific unit
291
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
292
+ Treatment of failed trials
293
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
294
+ Reported quantity
295
+ \end{minipage} \\
296
+ \midrule\noalign{}
297
+ Completion & Process, then repeated configuration & Retained as unsuccessful & Fraction completing the requested terminal step \\
298
+ Operational safety & Repeated configuration & Any eligible memory failure makes the configuration unsafe & All repetitions complete and the maximum peak is at most 38,912 MiB \\
299
+ Boundary position & Model--workload case & Unsafe settings remain in the comparison & Highest observed safe configuration level \\
300
+ Directional portability & Ordered source--target workload, model, or GPU allocation & Target failures and margin violations enter separate confusion counts & Approval errors and rejected within-margin settings \\
301
+ State-release effect & Base configuration & Pairs where completion differs are retained & Completion difference and peak difference where both finish \\
302
+ Longer-run validity & Screened configuration within a case & Failures remain in the comparison; failure after an accepted screen is a false-safe decision & One-step-to-long-run label transfer and later peak growth \\
303
+ \bottomrule
304
+ \end{tabular}
305
+ \end{table*}
306
+
307
+ Peak-memory comparisons require particular care. Whole-run contrasts use conditions that complete: a failed run's observed maximum is not its completed-run peak. A stage-specific contrast may include a later-failing run when the required stage finishes in both conditions. If failure interrupts that stage, its contrast is missing. All eligible trials remain in the completion and safety analyses, with contributing counts reported separately for continuous measurements.
308
+
309
+ \section{Experimental protocol}\label{experimental-protocol}
310
+
311
+ \subsection{Hardware and software}\label{hardware-and-software}
312
+
313
+ The experiments ran on Karolina, a supercomputer operated by IT4Innovations National Supercomputing Center in Ostrava, Czech Republic \cite{it4i2026karolina}. Its accelerated nodes contain eight NVIDIA A100 GPUs; our jobs used allocations of two or four A100-SXM4-40GB devices. Each device reported 40,960 MiB of memory, with driver version 610.43.02.
314
+
315
+ The container environment used Python 3.12.3, PyTorch 2.10.0 for model
316
+ computation, NVIDIA's CUDA execution platform version 12.9, and vLLM 0.18.0
317
+ for generation. The VERL source includes measurement instrumentation and
318
+ documented compatibility repairs. Thus, the execution design is held fixed,
319
+ but the source is not an unmodified upstream release. Exact source revisions,
320
+ all package versions, and the container checksum are retained with each
321
+ trial in the reproduction package.
322
+
323
+ \subsection{Training setup and configuration levels}\label{training-setup-and-configuration-levels}
324
+
325
+ All reported runs use the standard GRPO estimator \cite{shao2024deepseekmath}
326
+ with LoRA adapters \cite{hu2022lora}. The primary allocation places one vLLM
327
+ generation worker on each of two GPUs; a worker does not split its model
328
+ across GPUs. Fully Sharded Data Parallel (FSDP) partitions actor and
329
+ reference-policy state among the training workers across the allocation.
330
+
331
+ The boundary study uses Qwen2.5-3B-Instruct \cite{qwen2024qwen253b},
332
+ Phi-4-mini-instruct \cite{microsoft2025phi4mini}, and
333
+ Granite-3.3-2B-Instruct \cite{ibm2025granite33}, with approximately 3.1,
334
+ 3.8, and two billion parameters, respectively. State-release and original
335
+ temporal tests also use Qwen2.5-1.5B-Instruct. Tables omit the
336
+ instruction-tuned suffix; prose uses Qwen, Phi, and Granite when the
337
+ size is unambiguous.
338
+
339
+ Table 3 lists the fixed training controls. Model parameters and generation use the 16-bit bfloat16 format (BF16), while gradient reduction uses 32-bit floating point (FP32). Gradient checkpointing recomputes intermediate activations during training to reduce stored state. Kullback--Leibler (KL) divergence penalizes departure from the reference policy; it is applied as an actor loss rather than added to the reward.
340
+
341
+ AdamW is adaptive moment estimation with decoupled weight decay; its beta
342
+ parameters control running averages of gradients and their squares.
343
+ LoRA's rank sets the adapter dimension, and alpha scales its update.
344
+ Generation samples from token probabilities, whereas greedy validation
345
+ selects the most probable token. Temperature rescales these probabilities;
346
+ top-p and top-k restrict the probability mass and number of candidates,
347
+ respectively. The chosen values leave training sampling unrestricted.
348
+
349
+ Adapters are attached to the model's linear transformations. Padding removal avoids computing on padding tokens introduced to align variable-length sequences. Policy clipping limits the probability-ratio change used in the update objective, and gradient clipping bounds the gradient norm. The KL loss uses VERL's low-variance estimate of divergence. None of these controls is varied in the experiments.
350
+
351
+ \begin{table*}[!t]
352
+ \centering
353
+ \caption{Fixed training controls in the retained LoRA GRPO experiments.}
354
+ \footnotesize
355
+ \setlength{\tabcolsep}{3pt}
356
+ \renewcommand{\arraystretch}{1.12}
357
+ \begin{tabular}{@{}
358
+ >{\raggedright\arraybackslash}p{(\linewidth - 2\tabcolsep) * \real{0.5000}}
359
+ >{\raggedright\arraybackslash}p{(\linewidth - 2\tabcolsep) * \real{0.5000}}@{}}
360
+ \toprule\noalign{}
361
+ \begin{minipage}[b]{\linewidth}\raggedright
362
+ Fixed training control
363
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
364
+ Setting
365
+ \end{minipage} \\
366
+ \midrule\noalign{}
367
+ LoRA & Rank 64; alpha 32; all linear transformations \\
368
+ Gradient checkpointing & Enabled \\
369
+ Padding removal & Enabled \\
370
+ Actor FSDP mixed precision & BF16 parameters; FP32 gradient reduction \\
371
+ vLLM inference precision & BF16 \\
372
+ Optimizer & PyTorch AdamW; learning rate \(3 \times 10^{-6}\) \\
373
+ Optimizer controls & Betas (0.9, 0.999); weight decay 0.01; gradient clip 1.0 \\
374
+ Update schedule & Constant learning rate; no initial ramp; one pass over the update batch \\
375
+ Policy clipping / micro-batch sizing & Clip ratio 0.2 / fixed sizes \\
376
+ KL regularization & Actor loss enabled; coefficient 0.001; low-variance estimator \\
377
+ KL in reward / entropy bonus & Disabled / no bonus encouraging varied token choices \\
378
+ Data order & Fixed order; overlong prompts filtered \\
379
+ Reference parameter offload & Enabled \\
380
+ Training generation & Sampling enabled; temperature 1; top-p 1; top-k disabled \\
381
+ Validation generation & Greedy decoding; one response per prompt \\
382
+ \bottomrule
383
+ \end{tabular}
384
+ \end{table*}
385
+
386
+ Table 4 gives the workloads used for the repeated boundary and batch/reservation experiments. Prompt and response lengths are upper limits; individual samples may be shorter. The log-probability micro-batch applies to policy and reference log-probability scoring.
387
+
388
+ \begin{table*}[!t]
389
+ \centering
390
+ \caption{Workload settings for boundary and batch/reservation experiments.}
391
+ \footnotesize
392
+ \setlength{\tabcolsep}{3pt}
393
+ \renewcommand{\arraystretch}{1.12}
394
+ \begin{tabular}{@{}
395
+ >{\raggedright\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1765}}
396
+ >{\raggedleft\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1176}}
397
+ >{\raggedleft\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1176}}
398
+ >{\raggedleft\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1176}}
399
+ >{\raggedleft\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1176}}
400
+ >{\raggedleft\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1176}}
401
+ >{\raggedleft\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1176}}
402
+ >{\raggedleft\arraybackslash}p{(\linewidth - 14\tabcolsep) * \real{0.1176}}@{}}
403
+ \toprule\noalign{}
404
+ \begin{minipage}[b]{\linewidth}\raggedright
405
+ Workload
406
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
407
+ Prompt cap
408
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
409
+ Response cap
410
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
411
+ Model-length cap
412
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
413
+ Max sequences
414
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
415
+ Rollouts per prompt
416
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
417
+ Global batch
418
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
419
+ Log-prob. micro-batch
420
+ \end{minipage} \\
421
+ \midrule\noalign{}
422
+ GSM8K & 1,024 & 2,048 & 3,072 & 64 & 2 & 32 & 4 \\
423
+ MATH & 512 & 512 & 1,024 & 128 & 2 & 64 & 8 \\
424
+ Standard code & 640 & 1,024 & 1,664 & 64 & 2 & 32 & 4 \\
425
+ Longer-prompt code & 1,536 & 1,024 & 2,560 & 64 & 2 & 32 & 4 \\
426
+ \bottomrule
427
+ \end{tabular}
428
+ \end{table*}
429
+
430
+ The model-length cap bounds the combined prompt and response; the sequence cap bounds the number of simultaneously active generation sequences. Each boundary and batch/reservation process performs one training iteration. Section IV-C reports preparation-subset sizes and filtering. Other experiments use the settings specified in their respective protocols below.
431
+
432
+ \begin{table}[!t]
433
+ \centering
434
+ \caption{Six ordered configuration levels, from lower to higher memory pressure.}
435
+ \footnotesize
436
+ \setlength{\tabcolsep}{3pt}
437
+ \renewcommand{\arraystretch}{1.12}
438
+ \begin{tabular}{@{}
439
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2143}}
440
+ >{\raggedleft\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2857}}
441
+ >{\raggedleft\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2857}}
442
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2143}}@{}}
443
+ \toprule\noalign{}
444
+ \begin{minipage}[b]{\linewidth}\raggedright
445
+ Level
446
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
447
+ Actor micro-batch per GPU
448
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
449
+ vLLM reservation
450
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
451
+ Actor parameter offload
452
+ \end{minipage} \\
453
+ \midrule\noalign{}
454
+ c0 & 1 & 0.30 & Enabled \\
455
+ c1 & 4 & 0.45 & Enabled \\
456
+ c2 & 8 & 0.60 & Disabled \\
457
+ c3 & 16 & 0.70 & Disabled \\
458
+ c4 & 16 & 0.80 & Disabled \\
459
+ c5 & 32 & 0.85 & Disabled \\
460
+ \bottomrule
461
+ \end{tabular}
462
+ \end{table}
463
+
464
+ Table 5 defines the six configuration levels, c0--c5. They jointly increase actor micro-batch size and generation-memory reservation, with actor parameter offload disabled above c1. They are ordered test settings, not equally spaced values of one variable, so the comparison cannot isolate the effect of an individual control. Optimizer offload is disabled and generation-state release is enabled throughout.
465
+
466
+ \subsection{Experimental coverage}\label{experimental-coverage}
467
+
468
+ \begin{table*}[!t]
469
+ \centering
470
+ \caption{Recorded attempts for the original 588 planned configuration--seed slots. Completed runs and memory failures are the retained outcomes; excluded attempts are additional records.}
471
+ \footnotesize
472
+ \setlength{\tabcolsep}{3pt}
473
+ \renewcommand{\arraystretch}{1.12}
474
+ \begin{tabular}{@{}lrrrrr@{}}
475
+ \toprule\noalign{}
476
+ Study & Planned slots & Recorded attempts & Completed & Memory failures & Excluded \\
477
+ \midrule\noalign{}
478
+ Three-model boundary & 216 & 251 & 156 & 60 & 35 \\
479
+ Paired state release & 36 & 41 & 27 & 9 & 5 \\
480
+ Batch/reservation & 144 & 161 & 144 & 0 & 17 \\
481
+ Original 40-step runs & 18 & 18 & 15 & 3 & 0 \\
482
+ Original 100-step runs & 12 & 14 & 12 & 0 & 2 \\
483
+ Sampling calibration & 12 & 12 & 12 & 0 & 0 \\
484
+ Allocator-instrumentation control & 72 & 76 & 60 & 12 & 4 \\
485
+ Paired 100-step follow-up & 24 & 27 & 24 & 0 & 3 \\
486
+ Four-GPU transfer & 54 & 57 & 54 & 0 & 3 \\
487
+ \textbf{Total} & \textbf{588} & \textbf{657} & \textbf{504} & \textbf{84} & \textbf{69} \\
488
+ \bottomrule
489
+ \end{tabular}
490
+ \end{table*}
491
+
492
+ The original evidence comprises 438 initial and 150 follow-up retained outcomes. Table 6 counts recorded runner invocations, not queued jobs that never reached execution. The 69 exclusions comprise two completed processes belonging to invalid pairs, one incomplete memory-failure record, 55 other recorded failures, and eleven attempts without a trial record. The supplementary attempt inventory gives the reason and available evidence for every exclusion. Sixteen supporting historical screens and observations from the abandoned optimization objective remain separate. The targeted batch-by-logging control is also additional to Table 6; Section V-G reports its own attempt history.
493
+
494
+ \begin{table*}[!t]
495
+ \centering
496
+ \caption{Realized work in the boundary experiment. Lengths are tokens; means give equal weight to completed processes. ``Cap batches'' counts completed batches with at least one response reaching its configured limit, not the fraction of individual responses.}
497
+ \footnotesize
498
+ \setlength{\tabcolsep}{3pt}
499
+ \renewcommand{\arraystretch}{1.12}
500
+ \begin{tabular}{@{}llrrrrr@{}}
501
+ \toprule\noalign{}
502
+ Model & Workload & \shortstack{Examples before/\\after filtering} &
503
+ Completed & \shortstack{Prompt\\mean / max.} &
504
+ \shortstack{Response\\mean / max.} & Cap batches \\
505
+ \midrule\noalign{}
506
+ Qwen2.5-3B & GSM8K & 32 / 32 & 12 & 108 / 154 & 298 / 830 & 0 / 12 \\
507
+ Qwen2.5-3B & MATH & 64 / 64 & 12 & 104 / 437 & 392 / 512 & 12 / 12 \\
508
+ Qwen2.5-3B & Standard code & 32 / 32 & 12 & 421 / 518 & 286 / 958 & 0 / 12 \\
509
+ Qwen2.5-3B & Longer-prompt code & 32 / 32 & 12 & 720 / 772 & 352 / 1,024 & 7 / 12 \\
510
+ Phi-4-mini & GSM8K & 32 / 32 & 12 & 78 / 121 & 226 / 2,048 & 2 / 12 \\
511
+ Phi-4-mini & MATH & 64 / 64 & 12 & 76 / 401 & 367 / 512 & 12 / 12 \\
512
+ Phi-4-mini & Standard code & 32 / 32 & 12 & 379 / 480 & 326 / 1,024 & 12 / 12 \\
513
+ Phi-4-mini & Longer-prompt code & 32 / 32 & 12 & 664 / 746 & 412 / 1,024 & 12 / 12 \\
514
+ Granite-3.3-2B & GSM8K & 128 / 128 & 15 & 146 / 198 & 285 / 2,048 & 5 / 15 \\
515
+ Granite-3.3-2B & MATH & 256 / 252 & 15 & 132 / 460 & 376 / 512 & 15 / 15 \\
516
+ Granite-3.3-2B & Standard code & 64 / 47 & 15 & 492 / 588 & 228 / 1,024 & 1 / 15 \\
517
+ Granite-3.3-2B & Longer-prompt code & 64 / 63 & 15 & 807 / 902 & 296 / 1,024 & 1 / 15 \\
518
+ \bottomrule
519
+ \end{tabular}
520
+ \end{table*}
521
+
522
+ Table 7 reports the realized work in completed boundary runs. Each executes one global training batch. The larger Granite preparation subsets do not increase that batch, but model-specific filtering can change which examples reach it. All 156 completed runs have batch-length summaries; the 60 failed runs do not. Missing summaries are not interpreted as zero-length work. Machine-readable summaries cover all original studies; the supplement presents the completed 100-step studies and the new control, including fixed-order data reuse.
523
+
524
+ \subsection{Transfer between models}\label{transfer-between-models}
525
+
526
+ Each model is evaluated on all four workloads at the six levels in Table 5. Development runs checked that each case included both within-margin and unsuitable settings. Qwen and Phi use seeds 41--43 for the mathematics workloads and 51--53 for code, giving 144 trials. Granite adds 72 trials on the same workload and configuration grid.
527
+
528
+ The outcomes are repeatability of labels, the highest observed safe level, and directional transfer errors. Peak-memory differences use seed-matched runs in which both models complete. Because the model choices differ in architecture, parameter count, and tokenizer, these comparisons evaluate transfer between practical model choices rather than isolating any one of those properties.
529
+
530
+ The retrospective lookup baseline uses one donor workload's complete six-level profile to predict the other three workloads for the same model. Predictions distinguish within-margin completion, above-margin completion, and memory failure. Each donor profile requires eighteen retained processes; target outcomes are used only for scoring. We evaluate every donor choice, while recognizing that these comparisons reuse the same target configurations. The source-model baseline instead copies the label at a matched workload and level from another model. Approval errors are separated into actual memory failures and margin violations.
531
+
532
+ \subsection{Repetition policy}\label{repetition-policy}
533
+
534
+ Each boundary configuration, release/resident setting, batch/reservation setting, and long-run regime uses three fresh processes with fixed seeds. Seeds vary generation and execution randomness; they do not sample new workloads. The sampling-rate calibration instead uses six matched settings because it tests measurement agreement. Assignments within each experiment are fixed before its execution; the later choice of control cases using earlier results is documented separately.
535
+
536
+ The original longer-run study refers to 16 earlier one-step screens across six configurations. Five have three processes each; the unsafe MATH configuration has one. These supporting screens are audited separately and are not included in the principal count. Their logs and memory peaks must agree with the source values recorded before the longer runs.
537
+
538
+ The early source traces have stage labels but a constant zero step counter. Their one-step duration is verified from the configured limit and execution log, without changing those traces. Later-step growth is calculated only from traces that record step numbers. The follow-up study uses three fresh processes for both the short and long conditions.
539
+
540
+ \subsection{Testing one-step screens over longer runs}\label{testing-one-step-screens-over-longer-runs}
541
+
542
+ One-step runs make broad configuration screening affordable. To test their usefulness over time, the original longer-run experiment uses two model--workload cases and three initial regimes: comfortably below the operating threshold, near it, and unsafe in the available historical screen. Three seeds per setting give eighteen runs of forty requested steps. A preliminary timing check confirmed that the instrumentation recorded distinct training steps.
543
+
544
+ These cases use different workload settings from Table 4: Qwen2.5-3B/GSM8K has prompt/response caps 256/1,024 and model-length cap 1,280; Qwen2.5-1.5B/MATH has caps 1,792/1,024 and model-length cap 2,816. Both use global batch 128, five rollouts per prompt, at most 256 active sequences, log-probability micro-batch 8, untruncated training data, and up to 128 validation examples. Safe-margin and near-boundary reservations are 0.50 and 0.65. Their actor micro-batches are 16 for Qwen2.5-3B and 4 for Qwen2.5-1.5B; the unsafe regime uses reservation 0.80 and actor micro-batch 16. These are selected source regimes, not extra levels of the six-level boundary experiment.
545
+
546
+ The outcomes are completion, agreement with the one-step safety label, and the largest later increase over the memory peak at step 1. Reward and optimization statistics are retained as checks that training executed.
547
+
548
+ Actor parameter and optimizer offload are disabled, generation-state release is enabled, and the original forty-step runs have no validation or checkpoint writing. The historical label comparison uses the available source screens, including the unsafe MATH regime's single-screen exception; it is not a new three-seed screen for every regime.
549
+
550
+ \subsection{Memory-mechanism and 100-step experiments}\label{memory-mechanism-and-100-step-experiments}
551
+
552
+ The release experiment compares removing generation state with keeping it resident. Each pair runs both conditions sequentially on the same node and GPUs. Their order is counterbalanced: half run release first and half resident first. This is a \emph{crossover design}, in which each allocation receives both conditions. Before each period, the runner requires combined idle memory below 2,048 MiB across the two GPUs and records their baselines. Eligibility additionally requires records for both runs and identical GPU identities. This aggregate idle check differs from the later controls' per-GPU check.
553
+
554
+ Release enables vLLM Sleep Mode level 1 when replicas are parked after initialization or generation: generation weights are offloaded to host memory and the KV cache is discarded. The measured LoRA implementation restores weights before synchronization, transfers the base weights when first needed and subsequent adapter updates, then restores KV-cache capacity before generation resumes. Resident mode disables this sleep/wake cycle. It is therefore a state-lifecycle intervention, not merely a call to empty unused allocator cache. Both conditions use unmerged LoRA adapters, the same weight-transfer backend, and one training iteration without validation or checkpoints. Section V-C specifies all six base configurations and their outcomes; three seeds are used for each.
555
+
556
+ The original 100-step experiment inherits the forty-step study's two model--workload cases and the complete settings of its safe-margin and near-boundary regimes, including their offload and release controls. It uses three new seeds (91--93). Validation runs before training and every twenty steps; checkpoints are written at steps eighty and one hundred. Separate markers identify these operations. In addition to completion and safety, we compare their maximum memory use with the maximum over all other stages.
557
+
558
+ A separate \emph{factorial experiment} tests every combination of two factors: actor micro-batch size (8 or 16) and vLLM reservation (0.60 or 0.70). It covers three models, four workloads, and three common seeds. A four-condition block consists of the runs sharing a model, workload, and seed. The main effect of a factor averages its difference over the two values of the other factor; their interaction measures whether that difference depends on the other factor.
559
+
560
+ These original blocks are analytical groupings, not same-allocation crossovers. Their 144 processes ran in separate scheduler allocations across twenty nodes during the same experimental session. Placement and execution order were not randomized within blocks. Environment records identify each allocation and source state, including modified working copies. Thus, fixed scientific settings do not by themselves eliminate placement, order, or source-history effects. The new control below repeats the batch contrast in immutable, same-GPU blocks.
561
+
562
+ Safety effects use all blocks. A pairwise peak contrast requires both runs to complete, while a peak main effect or interaction requires all four peaks. The separate stage-specific reanalysis uses the same blocks and is identified as post hoc.
563
+
564
+ \subsection{Follow-up controls and transfer tests}\label{follow-up-controls-and-transfer-tests}
565
+
566
+ The instrumentation control tests whether allocator logging changes the measurement. It covers the three models on GSM8K and MATH at each model's previous highest safe and next unsafe level, using seeds 111--113. Each of 36 allocations runs two fresh processes on the same GPUs. The \emph{external-only} condition retains NVML sampling and stage markers. The \emph{full} condition additionally synchronizes GPU work before each traced worker operation, resets allocator peak counters, and records allocator statistics when the operation returns. The wrapper adds no separate synchronization after that return. Waiting for outstanding GPU operations can affect timing and stage attribution.
567
+
568
+ The two orders each occur in 18 allocations. Each period starts only after all GPUs return below 1,024 MiB of idle memory use. Completion and safety disagreements are primary outcomes; peak and elapsed-time differences use only pairs in which both conditions complete. All twelve base configurations are reported. This comparison measures the additional effect of allocator instrumentation, not the cost of all measurement activity.
569
+
570
+ The new batch-by-logging control uses Phi/GSM8K, where the existing stage measurements vary most, and Qwen2.5-3B/longer-prompt code, which adds code coverage. Each of three seeds (121--123) receives four fresh processes on the same two GPUs: micro-batch 8 or 16, each with external-only or full logging. Reservation is fixed at 0.60 and other workload settings follow Table 4. The six blocks use fixed counterbalanced condition orders, idle-device checks, an immutable source checkout, and no validation or checkpoints. The actor-stage batch effect subtracts the batch-8 peak from the batch-16 peak; the logging interaction subtracts the external-only batch effect from the full-logging effect. These are the primary outcomes. A pre-execution 512 MiB reference flags potentially material differences, not a certified equivalence margin. Every block and any incomplete attempt are retained.
571
+
572
+ The longer-run follow-up adds Phi and Granite on GSM8K and MATH with seeds 111--113. Each of twelve allocations runs a fresh one-step screen and a separate 100-step process at actor micro-batch 16. Reservations 0.65 (Phi) and 0.75 (Granite) interpolate between the previously observed safe/unsafe settings; they are fixed before new outcomes. Both horizons use untruncated training datasets, fixed order, validation before training and every 20 steps, and checkpoint frequency 25. Their GSM8K/MATH shapes match Table 4, with validation subsets of 32/64 examples respectively. The final step also triggers validation and checkpoint writing, including in the one-step screen. Only the latest saved state is kept, but all measurement events remain. Runs start without resuming a checkpoint, and long runs are kept even when the paired screen is unsafe. Validation checks require all 100 steps, evaluations at 0, 20, 40, 60, 80, and 100, and checkpoints at 25, 50, 75, and 100. The primary comparison uses the three-seed safety rule; individual paired outcomes are reported separately.
573
+
574
+ The four-GPU study retains one generation worker per GPU without splitting an individual worker's model. Three models, two mathematics workloads, levels c2--c4, and seeds 111--113 give 54 processes. Global batch, rollout count, and the remaining settings are unchanged. We compare its labels in both directions with the earlier two-GPU results. Because GPU count was not randomized within a common experimental period, this is a transfer test rather than an isolated causal estimate of adding GPUs.
575
+
576
+ \subsection{Statistics}\label{statistics}
577
+
578
+ We use a \emph{cluster bootstrap} to describe sensitivity to the composition of the measured case set, resampling groups of related observations rather than individual processes. The groups are six base configurations for state release, four workloads for model and factorial comparisons, and twelve base configurations for the original instrumentation control. Each nominal 95\% confidence interval (CI) uses 20,000 resamples with recorded seeds. The three base configurations where both release conditions complete contribute to its peak contrast.
579
+
580
+ These selected groups are neither numerous nor independent samples of a deployment population. We therefore emphasize per-case outcomes, effect ranges, and agreement in direction. A sensitivity analysis gives the two code subsets half weight each, so that the three source corpora contribute equally. The two-case original temporal study and the new two-case control are reported by case without population-risk bounds or bootstrap precision claims.
581
+
582
+ The post hoc stage-specific intervals are descriptive and have no adjustment for testing multiple effects. They are not evidence from a prospectively registered hypothesis test.
583
+
584
+ \section{Results}\label{results}
585
+
586
+ \subsection{Completion and memory safety}\label{completion-and-memory-safety}
587
+
588
+ The three-model boundary experiment contains 216 processes representing 72 configurations. Table 8 distinguishes completion from safety: 156 processes complete, but only 132 leave the specified margin. The sixty failures occur during generation initialization (51), generation itself (six), or weight synchronization (three).
589
+
590
+ All three repetitions agree within each configuration. There are 44 safe configurations and 28 unsafe configurations; the latter include both memory failures and runs that complete above the operating limit.
591
+
592
+ \begin{table*}[!t]
593
+ \centering
594
+ \caption{Completion and safety in the original boundary experiment.}
595
+ \footnotesize
596
+ \setlength{\tabcolsep}{3pt}
597
+ \renewcommand{\arraystretch}{1.12}
598
+ \begin{tabular}{@{}
599
+ >{\raggedright\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1304}}
600
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1739}}
601
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1739}}
602
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1739}}
603
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1739}}
604
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1739}}@{}}
605
+ \toprule\noalign{}
606
+ \begin{minipage}[b]{\linewidth}\raggedright
607
+ Model
608
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
609
+ Processes
610
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
611
+ Completed
612
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
613
+ Within margin
614
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
615
+ Memory failures
616
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
617
+ Safe configurations
618
+ \end{minipage} \\
619
+ \midrule\noalign{}
620
+ Qwen2.5-3B & 72 & 48 & 48 & 24 & 16/24 \\
621
+ Phi-4-mini & 72 & 48 & 36 & 24 & 12/24 \\
622
+ Granite-3.3-2B & 72 & 60 & 48 & 12 & 16/24 \\
623
+ \textbf{Total} & \textbf{216} & \textbf{156} & \textbf{132} & \textbf{60} & \textbf{44/72} \\
624
+ \bottomrule
625
+ \end{tabular}
626
+ \end{table*}
627
+
628
+ Figure 1 shows where each outcome occurs within the tested configuration levels.
629
+
630
+ \begin{figure*}[!t]
631
+ \centering
632
+ \pandocbounded{\includegraphics[keepaspectratio,alt={Three-state feasibility profiles across models, workloads, and configuration levels. Profiles are invariant across the four workloads within each model.}]{figures/standard_grpo_boundary.pdf}}
633
+ \caption{Three-state feasibility profiles across the boundary grid. Each cell combines three fresh processes. Identical profiles within a model explain the success of the donor-workload lookup baseline.}
634
+ \end{figure*}
635
+
636
+ \subsection{Simple lookup and cross-model transfer}\label{model-labels-are-not-interchangeable}
637
+
638
+ Within each model, the full six-level, three-state profile is identical across the four workloads. A donor-workload lookup therefore predicts every remaining workload label correctly. Across twelve donor choices, this gives 216 correct comparisons involving 72 distinct target configurations, not 216 independent observations. This grid does not demonstrate a need for a learned workload-specific predictor.
639
+
640
+ The model-specific profiles differ. Qwen and Granite remain within margin through c3, whereas Phi does so through c2. Table 9 shows the consequences of copying these profiles between models.
641
+
642
+ \begin{table*}[!t]
643
+ \centering
644
+ \caption{Source-model lookup at the primary margin. Each direction compares 24 target configurations. Approval errors distinguish actual memory failure from completion above margin; rejected usable settings are reported separately.}
645
+ \footnotesize
646
+ \setlength{\tabcolsep}{3pt}
647
+ \renewcommand{\arraystretch}{1.12}
648
+ \begin{tabular}{@{}lrrrrr@{}}
649
+ \toprule\noalign{}
650
+ Source $\rightarrow$ target & \shortstack{Correct\\three-state label} &
651
+ Approved & \shortstack{Approved:\\memory failure} &
652
+ \shortstack{Approved:\\above margin} & \shortstack{Rejected:\\within margin} \\
653
+ \midrule\noalign{}
654
+ Qwen $\rightarrow$ Phi & 20 & 16 & 0 & 4 & 0 \\
655
+ Granite $\rightarrow$ Phi & 16 & 16 & 0 & 4 & 0 \\
656
+ Phi $\rightarrow$ Qwen & 20 & 12 & 0 & 0 & 4 \\
657
+ Phi $\rightarrow$ Granite & 16 & 12 & 0 & 0 & 4 \\
658
+ Qwen $\rightarrow$ Granite & 20 & 16 & 0 & 0 & 0 \\
659
+ Granite $\rightarrow$ Qwen & 20 & 16 & 0 & 0 & 0 \\
660
+ \bottomrule
661
+ \end{tabular}
662
+ \end{table*}
663
+
664
+ Transferring Qwen's labels to Phi accepts four of Phi's twelve unsafe configurations. All four complete above the operating limit: these errors concern the reserved memory margin, not missed OOMs. Qwen and Granite agree on binary safety at all 24 settings, although their three-state profiles differ because Granite completes above margin at c4, where Qwen fails.
665
+
666
+ The four Qwen-to-Phi errors are also four of the 16 configurations approved by Qwen. Thus, the false-safe fraction among unsafe targets and the error fraction among approved settings answer different questions. Neither denominator represents a sampled deployment workload.
667
+
668
+ At the sixteen matched settings where both models complete, Phi uses an average of 3,517 MiB more peak memory than Qwen (95\% CI 3,472 to 3,562 MiB) and 4,598 MiB more than Granite (95\% CI 4,480 to 4,687 MiB). Granite uses 1,081 MiB less than Qwen (95\% CI 976 to 1,144 MiB less). These intervals describe variation under resampling of the four selected workloads, not precision about unseen workloads. The Qwen/Phi boundary difference has the same direction in all four.
669
+
670
+ The distinction depends on the reserved margin. Table 10 recalculates the labels under five operating thresholds; memory failures remain unsafe throughout. This sensitivity analysis was performed after the primary experiment.
671
+
672
+ \begin{table*}[!t]
673
+ \centering
674
+ \caption{Post hoc sensitivity to the reserved memory margin.}
675
+ \footnotesize
676
+ \setlength{\tabcolsep}{3pt}
677
+ \renewcommand{\arraystretch}{1.12}
678
+ \begin{tabular}{@{}
679
+ >{\raggedright\arraybackslash}p{(\linewidth - 8\tabcolsep) * \real{0.1765}}
680
+ >{\raggedleft\arraybackslash}p{(\linewidth - 8\tabcolsep) * \real{0.2353}}
681
+ >{\raggedright\arraybackslash}p{(\linewidth - 8\tabcolsep) * \real{0.1765}}
682
+ >{\raggedright\arraybackslash}p{(\linewidth - 8\tabcolsep) * \real{0.1765}}
683
+ >{\raggedleft\arraybackslash}p{(\linewidth - 8\tabcolsep) * \real{0.2353}}@{}}
684
+ \toprule\noalign{}
685
+ \begin{minipage}[b]{\linewidth}\raggedright
686
+ Capacity allowed
687
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
688
+ Limit (MiB)
689
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
690
+ Qwen highest safe level
691
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
692
+ Phi highest safe level
693
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
694
+ Qwen→Phi false-safe / unsafe target configurations
695
+ \end{minipage} \\
696
+ \midrule\noalign{}
697
+ 90\% & 36,864 & c2, all workloads & c2, all workloads & 0/12 \\
698
+ 92.5\% & 37,888 & c3, all workloads & c2, all workloads & 4/12 \\
699
+ 95\% (primary) & 38,912 & c3, all workloads & c2, all workloads & 4/12 \\
700
+ 97.5\% & 39,936 & c3, all workloads & c3 GSM8K; c2 otherwise & 3/11 \\
701
+ 100\% & 40,960 & c3, all workloads & c3, all workloads & 0/8 \\
702
+ \bottomrule
703
+ \end{tabular}
704
+ \end{table*}
705
+
706
+ Granite reaches c3 at the primary limit and c4 when no margin is reserved. At 90\% and 92.5\%, it reaches c3 except on MATH, where it reaches c2. Qwen and Phi have the same completion boundary at the full-capacity limit. The reproduction package includes all 60 model--workload--threshold combinations and 30 directional transfer summaries.
707
+
708
+ \subsection{Releasing generation state changes completion}\label{releasing-generation-state-changes-completion}
709
+
710
+ In the 18 paired experiments on the same GPUs, release succeeds and resident fails in nine pairs; both complete in the other nine. No pair completes only in resident mode. Table 11 gives the settings and outcomes, and Figure 2 summarizes the memory-mechanism experiments. The completion difference is 50.0 percentage points (95\% CI 16.7 to 83.3), with resampling by base configuration. Six resident-mode failures occur during weight synchronization and three during actor update.
711
+
712
+ \begin{table*}[!t]
713
+ \centering
714
+ \caption{All state-release base configurations. Prompt/response limits are tokens; their sum is the model-length limit. The selected training subset equals the global batch. Actor and log-probability micro-batches are per GPU; the latter applies to both policy and reference scoring. Completed counts are release / resident, each out of three.}
715
+ \footnotesize
716
+ \setlength{\tabcolsep}{3pt}
717
+ \renewcommand{\arraystretch}{1.12}
718
+ \begin{tabular}{@{}llrrrrrlr@{}}
719
+ \toprule\noalign{}
720
+ Model & Workload & \shortstack{Prompt /\\response cap} &
721
+ \shortstack{Max.\\sequences} & \shortstack{Rollouts /\\prompt} &
722
+ \shortstack{Global\\batch} & \shortstack{Actor / log-prob.\\micro-batch} &
723
+ \shortstack{Reservation;\\actor offload} & Completed \\
724
+ \midrule\noalign{}
725
+ Qwen2.5-1.5B & GSM8K & 1,024 / 2,048 & 64 & 2 & 32 & 16 / 8 & 0.70; no & 3 / 3 \\
726
+ Qwen2.5-1.5B & MATH & 512 / 512 & 128 & 2 & 64 & 16 / 8 & 0.70; no & 3 / 0 \\
727
+ Qwen2.5-3B & GSM8K & 1,024 / 2,048 & 64 & 2 & 32 & 1 / 4 & 0.30; yes & 3 / 3 \\
728
+ Qwen2.5-3B & GSM8K & 256 / 1,024 & 256 & 5 & 128 & 16 / 8 & 0.65; no & 3 / 0 \\
729
+ Qwen2.5-3B & MATH & 1,792 / 1,024 & 256 & 5 & 128 & 16 / 8 & 0.65; no & 3 / 0 \\
730
+ Qwen2.5-3B & MATH & 1,792 / 1,024 & 256 & 5 & 128 & 2 / 8 & 0.35; yes & 3 / 3 \\
731
+ \bottomrule
732
+ \end{tabular}
733
+ \end{table*}
734
+
735
+ Among the nine pairs where both conditions complete, retaining generation state increases peak memory by 7,061 MiB on average (95\% CI 5,249 to 9,391 MiB), with three contributing base configurations. Five release-only completions occur in the nine release-first allocations and four in the nine resident-first allocations. Thus, the completion difference appears in both orders. The experiment does not establish that release has no throughput or latency cost.
736
+
737
+ \begin{figure*}[!t]
738
+ \centering
739
+ \pandocbounded{\includegraphics[keepaspectratio,alt={Memory-mechanism experiments. Panel (a) shows paired release/resident outcomes by execution order. Panels (b) and (c) show safety and peak-memory effects from the batch/reservation experiment; its intervals resample the four workloads. Peak contrasts use completed runs.}]{figures/standard_grpo_mechanisms.pdf}}
740
+ \caption{Memory-mechanism experiments. Panel (a) shows paired release/resident outcomes by execution order. Panels (b) and (c) show safety and peak-memory effects from the batch/reservation experiment; its intervals resample the four workloads. Peak contrasts use completed runs.}
741
+ \end{figure*}
742
+
743
+ \subsection{The overall peak hides the training-batch effect}\label{the-overall-peak-hides-the-training-batch-effect}
744
+
745
+ All 144 processes in the batch/reservation experiment complete. Of these, 120 leave the specified margin and 24 cross the operating threshold. The three-repetition rule yields forty safe and eight unsafe configurations; none of the unsafe labels in this experiment is caused by a crash.
746
+
747
+ Increasing vLLM reservation from 0.60 to 0.70 raises peak memory by 3,998 MiB (95\% CI 3,918 to 4,052 MiB). The fraction of safe configurations falls by 33.3 percentage points in each workload: all Phi settings cross the limit, while Qwen and Granite retain their labels. The memory increase occurs across all three models.
748
+
749
+ Doubling actor micro-batch from 8 to 16 leaves the safety labels unchanged and changes the overall peak by only -31 MiB (95\% CI -75 to -3 MiB). The peak-memory interaction is +21 MiB (95\% CI -57 to +135 MiB); the safety interaction is zero in every workload.
750
+
751
+ The reservation effect is close to the 4,096 MiB implied by an additional 10\% of device capacity. The small overall batch effect, however, conceals a much larger increase during actor update (Table 12). A post hoc analysis of the 36 complete four-condition blocks under full instrumentation finds a 3,980 MiB increase in the actor-update NVML peak (95\% CI 2,588 to 5,563 MiB). PyTorch's allocated memory peak increases by 3,848 MiB (95\% CI 2,502 to 5,383 MiB), taking the maximum across the distributed workers. These magnitudes describe the instrumented execution; the targeted control evaluates their sensitivity to logging.
752
+
753
+ \begin{table*}[!t]
754
+ \centering
755
+ \caption{Overall and actor-update effects under full instrumentation. Brackets give 95\% bootstrap resampling intervals over the four selected workloads; the new controls assess magnitude sensitivity.}
756
+ \footnotesize
757
+ \setlength{\tabcolsep}{3pt}
758
+ \renewcommand{\arraystretch}{1.12}
759
+ \begin{tabular}{@{}
760
+ >{\raggedright\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2000}}
761
+ >{\raggedleft\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2667}}
762
+ >{\raggedleft\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2667}}
763
+ >{\raggedleft\arraybackslash}p{(\linewidth - 6\tabcolsep) * \real{0.2667}}@{}}
764
+ \toprule\noalign{}
765
+ \begin{minipage}[b]{\linewidth}\raggedright
766
+ Factorial main effect
767
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
768
+ Whole-run NVML peak (MiB)
769
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
770
+ Actor-update NVML peak (MiB)
771
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
772
+ Actor-update allocated peak (MiB)
773
+ \end{minipage} \\
774
+ \midrule\noalign{}
775
+ Actor micro-batch 8→16 & −31 {[}−75, −3{]} & 3,980 {[}2,588, 5,563{]} & 3,848 {[}2,502, 5,383{]} \\
776
+ Reservation 0.60→0.70 & 3,998 {[}3,918, 4,052{]} & 9 {[}−216, 170{]} & 1 {[}−200, 132{]} \\
777
+ \bottomrule
778
+ \end{tabular}
779
+ \end{table*}
780
+
781
+ Actor-update NVML effects differ across workloads: 2,073 MiB for GSM8K, 3,448 MiB for MATH, 4,132 MiB for standard code, and 6,268 MiB for longer-prompt code. Giving the two code subsets half weight each reduces the mean from 3,980 to 3,574 MiB; all workload effects remain positive. The reservation effect is essentially unchanged at 4,006 MiB under this weighting.
782
+
783
+ Actor update determines the overall peak in none of the 144 traces. Generation reaches it in 74 runs and weight synchronization in 47. The remaining peaks occur in intervals marked idle (19) or initialization (four); idle-only observations cannot be attributed to a specific operation. These records explain how substantial actor-stage growth can leave the whole-run maximum nearly unchanged. Figure 3 presents the targeted control, discussed in Section V-G, that tests whether the batch contrast survives without allocator instrumentation.
784
+
785
+ \begin{figure*}[!t]
786
+ \centering
787
+ \pandocbounded{\includegraphics[keepaspectratio,alt={Actor-stage batch-size effects under external-only and full instrumentation in the targeted control. Connected points share a seed and GPU allocation.}]{figures/review_instrumentation_batch.pdf}}
788
+ \caption{Targeted batch-by-logging control. Each connected pair compares the actor-stage batch-16 minus batch-8 effect under external-only and full instrumentation within one four-condition allocation. Asterisks identify additional allocations after initial noncompletion. These are observed block effects, not confidence intervals. Section V-G reports the attempt history.}
789
+ \end{figure*}
790
+
791
+ \subsection{Longer runs retain labels despite memory growth}\label{longer-runs-retain-labels-despite-memory-growth}
792
+
793
+ In the original forty-step study, all twelve runs from initially safe configurations complete and remain within margin. Of the six runs from unsafe configurations, three complete above the operating threshold and three fail during generation initialization. All eighteen labels agree with the available historical screens, including the single-screen exception described in the protocol. Among the fifteen completed runs, the maximum later increase over step 1 averages 1,717 MiB and reaches 2,404 MiB.
794
+
795
+ All twelve runs in the original 100-step extension also complete and remain safe. Their maximum later increase averages 1,925 MiB. Validation and checkpoint writing stay below the largest peak of the other stages by an average of 2,352 MiB; their respective mean peaks are 31,245 MiB and 16,708 MiB. Figure 4 shows the individual trajectories, and the supplement reports both case-level effects.
796
+
797
+ These are two selected model--workload cases. Their agreement does not estimate the risk of later failure in unseen settings.
798
+
799
+ \begin{figure*}[!t]
800
+ \centering
801
+ \pandocbounded{\includegraphics[keepaspectratio,alt={Step-1 and maximum memory peaks for each of the twelve original 100-step runs, separated by model and workload.}]{figures/standard_grpo_temporal.pdf}}
802
+ \caption{Individual trajectories in the original 100-step experiment. N and S denote the near-boundary and safe-margin regimes; the number identifies the repetition. Each line connects step 1 to the maximum within the same run.}
803
+ \end{figure*}
804
+
805
+ \subsection{Measurement calibration}\label{measurement-calibration}
806
+
807
+ All six paired runs comparing 10 ms and 100 ms sampling complete. Resampling the 10 ms traces at coarser intervals and different starting offsets gives 222 derived checks up to 200 ms, not additional independent runs. None changes the observed peak or its assigned stage. Runs sampled at 10 ms are 1.8 seconds shorter on average. These results support 100 ms sampling for the sustained peaks observed here; they cannot rule out missed shorter transients elsewhere.
808
+
809
+ \subsection{Effect of allocator instrumentation}\label{effect-of-allocator-instrumentation}
810
+
811
+ The instrumentation comparison contains 72 processes in 36 eligible pairs. Completion differs in 0/36 pairs, and safety differs in 0/36 pairs. Both conditions complete in 30/36 pairs. Table 13 reports each base configuration. Completed and safe counts are given for external-only and full instrumentation, respectively; each condition has three repetitions. Peak and time differences are full minus external-only and use only pairs where both complete.
812
+
813
+ \begin{table*}[!t]
814
+ \centering
815
+ \caption{Allocator-instrumentation control. Paired counts are external-only; full.}
816
+ \footnotesize
817
+ \setlength{\tabcolsep}{3pt}
818
+ \renewcommand{\arraystretch}{1.12}
819
+ \begin{tabular}{@{}
820
+ >{\raggedright\arraybackslash}p{(\linewidth - 12\tabcolsep) * \real{0.2667}}
821
+ >{\raggedright\arraybackslash}p{(\linewidth - 12\tabcolsep) * \real{0.0667}}
822
+ >{\raggedleft\arraybackslash}p{(\linewidth - 12\tabcolsep) * \real{0.1333}}
823
+ >{\raggedleft\arraybackslash}p{(\linewidth - 12\tabcolsep) * \real{0.1333}}
824
+ >{\raggedleft\arraybackslash}p{(\linewidth - 12\tabcolsep) * \real{0.1333}}
825
+ >{\raggedleft\arraybackslash}p{(\linewidth - 12\tabcolsep) * \real{0.1333}}
826
+ >{\raggedleft\arraybackslash}p{(\linewidth - 12\tabcolsep) * \real{0.1333}}@{}}
827
+ \toprule\noalign{}
828
+ \begin{minipage}[b]{\linewidth}\raggedright
829
+ Model / workload
830
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
831
+ Level
832
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
833
+ Completed
834
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
835
+ Safe
836
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
837
+ Label disagreements
838
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
839
+ Peak difference (MiB)
840
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
841
+ Time difference (s)
842
+ \end{minipage} \\
843
+ \midrule\noalign{}
844
+ Granite-3.3-2B / GSM8K & c3 & 3; 3 & 3; 3 & 0/3 & 0.0 & -15.3 \\
845
+ Granite-3.3-2B / GSM8K & c4 & 3; 3 & 0; 0 & 0/3 & 0.0 & -6.3 \\
846
+ Granite-3.3-2B / MATH & c3 & 3; 3 & 3; 3 & 0/3 & 0.0 & 4.7 \\
847
+ Granite-3.3-2B / MATH & c4 & 3; 3 & 0; 0 & 0/3 & 0.0 & -8.0 \\
848
+ Phi-4-mini / GSM8K & c2 & 3; 3 & 3; 3 & 0/3 & -6.7 & 11.0 \\
849
+ Phi-4-mini / GSM8K & c3 & 3; 3 & 0; 0 & 0/3 & -13.3 & -4.3 \\
850
+ Phi-4-mini / MATH & c2 & 3; 3 & 3; 3 & 0/3 & -149.3 & 5.0 \\
851
+ Phi-4-mini / MATH & c3 & 3; 3 & 0; 0 & 0/3 & -44.7 & 5.0 \\
852
+ Qwen2.5-3B / GSM8K & c3 & 3; 3 & 3; 3 & 0/3 & 0.0 & -19.3 \\
853
+ Qwen2.5-3B / GSM8K & c4 & 0; 0 & 0; 0 & 0/3 & --- & --- \\
854
+ Qwen2.5-3B / MATH & c3 & 3; 3 & 3; 3 & 0/3 & 0.0 & 13.7 \\
855
+ Qwen2.5-3B / MATH & c4 & 0; 0 & 0; 0 & 0/3 & --- & --- \\
856
+ \bottomrule
857
+ \end{tabular}
858
+ \end{table*}
859
+
860
+ There were forty external-only attempts and thirty-six full-instrumentation attempts. Two external attempts had diagnosed setup problems, and two reached the startup limit without a trial record. The latter remain unresolved; their full-instrumentation counterparts never ran. Excluding the two diagnosed setup attempts leaves 38 allocation attempts with between zero and two possible label disagreements, or 0--5.3\%, depending on the unknown outcomes. This is a missing-outcome sensitivity range, not an observed paired result or confidence interval. The zero disagreement count among eligible pairs does not establish equivalence.
861
+
862
+ Among completed pairs, the mean peak difference is -21.4 {[}-53.7, -1.3{]}
863
+ MiB. The elapsed-time difference is -1.4 {[}-8.0, 5.1{]} seconds, or
864
+ -0.5 {[}-3.4, 2.6{]}\%. Brackets give 95\% intervals from resampling the
865
+ twelve base configurations in the design; the continuous contrasts have
866
+ thirty contributing pairs from ten base configurations. A dash in Table 13
867
+ means neither condition completed; no missing peak is estimated.
868
+
869
+ Whole-run agreement is not sufficient to validate individual stages. Actor-update peaks are observed in both arms of 33 pairs, including three pairs that later fail. Median full-minus-external differences for the eleven available base configurations range from $-176$ to $400$ MiB. Individual differences range from $-1,800$ to $1,456$ MiB. The Qwen/MATH c4 pairs fail before actor update, so that stage has no comparison. The supplement gives every stage contrast. This variation motivated the direct batch-by-logging control rather than an assumption that whole-run agreement extends to actor-stage effects.
870
+
871
+ The targeted control initially completed four of its six allocations. Sixteen processes finished within margin; two full-logging processes, each first in its assigned order, did not finish within the twenty-minute limit. One logged an accelerator argument error during weight-transfer buffer reconstruction, while the other's cause remains unresolved. Neither is established as a memory-capacity failure. Their remaining six conditions never started. A dated amendment, recorded before any additional outcomes, allowed one further unchanged allocation for each incomplete block without a diagnosed memory failure. These additional attempts do not replace the initial noncompletion or add new seeds.
872
+
873
+ Both additional allocations completed, yielding eight further processes and six usable four-condition blocks overall. Thus, 24 of 26 attempted processes completed within margin. The actor-stage batch effect is positive under external-only measurement in every usable block: its case means are 1,966 MiB for Phi/GSM8K and 6,430 MiB for Qwen/longer-prompt code (Figure 3). The corresponding full-logging means are 2,287 and 7,059 MiB. The full-minus-external batch contrasts range from $-220$ to 1,054 MiB across blocks; three of six exceed the 512 MiB diagnostic in absolute value. Whole-run batch effects remain between $-160$ and 132 MiB.
874
+
875
+ These results support the direction of the actor-stage increase without allocator logging, but not an instrumentation-independent effect size. Prompt batches match within each case, whereas generated responses are not identical across conditions despite matched seeds. The contrasts therefore include execution and generation variability, not only tracing overhead on fixed tensors. Nor do the two initial full-first failures establish a logging effect: condition, order, seed, and allocation remain possible explanations. The supplement reports all block effects, realized response lengths, and initial and additional attempts separately.
876
+
877
+ \subsection{Longer-run validation on Phi and Granite}\label{longer-run-validation-on-phi-and-granite}
878
+
879
+ The follow-up compares twelve fresh one-step screens with separate 100-step runs (24 processes in total). All twelve long runs complete. However, two of the eleven individually safe screens are paired with long runs that cross the operating threshold. Both occur for Granite on MATH.
880
+
881
+ Table 14 separates individual completion and safety counts from the label obtained by requiring all three repetitions to be safe.
882
+
883
+ \begin{table*}[!t]
884
+ \centering
885
+ \caption{One-step and 100-step outcomes, with memory measurements from completed long runs.}
886
+ \footnotesize
887
+ \setlength{\tabcolsep}{3pt}
888
+ \renewcommand{\arraystretch}{1.12}
889
+ \begin{tabular}{@{}
890
+ >{\raggedright\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.2857}}
891
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1429}}
892
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1429}}
893
+ >{\raggedright\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1429}}
894
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1429}}
895
+ >{\raggedleft\arraybackslash}p{(\linewidth - 10\tabcolsep) * \real{0.1429}}@{}}
896
+ \toprule\noalign{}
897
+ \begin{minipage}[b]{\linewidth}\raggedright
898
+ Model / workload
899
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
900
+ Screen completed; safe
901
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
902
+ Long completed; safe
903
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedright
904
+ Repeated labels
905
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
906
+ Mean later growth (MiB)
907
+ \end{minipage} & \begin{minipage}[b]{\linewidth}\raggedleft
908
+ Mean validation / checkpoint peak (MiB)
909
+ \end{minipage} \\
910
+ \midrule\noalign{}
911
+ Granite-3.3-2B / GSM8K & 3/3; 3/3 & 3/3; 3/3 & safe → safe & 159 & 38,443 / 13,498 \\
912
+ Granite-3.3-2B / MATH & 3/3; 2/3 & 3/3; 0/3 & unsafe → unsafe & 657 & 38,875 / 15,039 \\
913
+ Phi-4-mini / GSM8K & 3/3; 3/3 & 3/3; 3/3 & safe → safe & 115 & 37,581 / 19,581 \\
914
+ Phi-4-mini / MATH & 3/3; 3/3 & 3/3; 3/3 & safe → safe & 590 & 37,883 / 24,620 \\
915
+ \bottomrule
916
+ \end{tabular}
917
+ \end{table*}
918
+
919
+ Under the three-seed rule, none of the three initially safe cases becomes unsafe. Granite--MATH was already rejected because one short run exceeded the limit. All long runs pass the required checks for training steps, validation, and checkpoint writing. Later growth is measured within each long run relative to its own step 1, not by subtracting the peak of the separate one-step screen.
920
+
921
+ \subsection{Transfer between GPU counts}\label{transfer-between-gpu-counts}
922
+
923
+ On four GPUs, all 54 processes complete and 48 leave the specified margin. The repeated labels therefore identify sixteen of eighteen configurations as safe.
924
+
925
+ \begin{table}[!t]
926
+ \centering
927
+ \caption{Outcomes at the same configuration levels on two and four GPUs.}
928
+ \footnotesize
929
+ \setlength{\tabcolsep}{3pt}
930
+ \renewcommand{\arraystretch}{1.12}
931
+ \begin{tabular}{@{}lll@{}}
932
+ \toprule\noalign{}
933
+ Model / workload & Two-GPU c2 / c3 / c4 & Four-GPU c2 / c3 / c4 \\
934
+ \midrule\noalign{}
935
+ Granite-3.3-2B / GSM8K & safe / safe / above & safe / safe / safe \\
936
+ Granite-3.3-2B / MATH & safe / safe / above & safe / safe / safe \\
937
+ Phi-4-mini / GSM8K & safe / above / failure & safe / safe / above \\
938
+ Phi-4-mini / MATH & safe / above / failure & safe / safe / above \\
939
+ Qwen2.5-3B / GSM8K & safe / safe / failure & safe / safe / safe \\
940
+ Qwen2.5-3B / MATH & safe / safe / failure & safe / safe / safe \\
941
+ \bottomrule
942
+ \end{tabular}
943
+ \end{table}
944
+
945
+ In Table 15, ``above'' denotes completion above the operating limit, and ``failure'' denotes at least one memory failure. Across the eighteen matched configurations, transferring two-GPU labels to four GPUs wrongly rejects six within-margin configurations and makes no approval error. Reverse transfer approves two configurations that fail and four that complete above margin. These six errors are six of eight unsafe two-GPU targets, or six of sixteen configurations approved on four GPUs. The tested c2--c4 range does not establish the four-GPU boundary beyond those levels.
946
+
947
+ \section{Discussion}\label{discussion}
948
+
949
+ \subsection{What can be transferred between configurations}\label{what-can-be-transferred-between-configurations}
950
+
951
+ The model and GPU-count experiments point to the same practical limitation: a safety label belongs to an execution setting, not to a batch size or memory-reservation value in isolation. Similar parameter counts do not ensure that two models have interchangeable boundaries. Likewise, a setting accepted on a larger allocation may be unsafe on a smaller one.
952
+
953
+ This matters when using the benchmark to guide configuration selection. A low average prediction error would not by itself demonstrate useful selection near the memory limit. An evaluation must show how often the chosen settings fail or exceed the margin, as well as how often usable settings are rejected.
954
+
955
+ The lookup results also identify a limitation of the benchmark grid. Within a model, workload-specific learning adds no value for the observed three-state labels because every workload has the same profile. A future selection method should therefore be compared against this simple rule, with donor measurements and retries included in its cost. Finer settings near the boundary or different workloads would be needed to test whether more complex prediction is worthwhile.
956
+
957
+ \subsection{Why stage-level measurements matter}\label{why-stage-level-measurements-matter}
958
+
959
+ The release and batch experiments reveal two different weaknesses of successful-run peak summaries. First, a summary restricted to completed runs cannot represent the settings that fail when generation state remains resident. Second, the overall maximum can hide a large increase in a stage that does not yet determine that maximum.
960
+
961
+ The latter distinction is important for tuning. An apparently negligible batch effect on the overall peak does not mean that increasing the batch has little memory cost. Another stage remains larger in the measured configuration; if that balance changes, the previously hidden cost can become limiting. The external-only controls support this interpretation while warning against treating the instrumented effect size as a universal correction.
962
+
963
+ \subsection{How to use a short-run screen}\label{how-to-use-a-short-run-screen}
964
+
965
+ The longer-run experiments distinguish three levels of evidence: one successful screen, agreement across repeated screens, and direct validation over a longer horizon. They should not be treated as interchangeable.
966
+
967
+ The Granite--MATH follow-up illustrates why. Two individual screens suggest safety that their paired long runs do not retain, but the three-seed rule already rejects the case because another short run exceeds the margin. Repeated screening therefore catches the problem in this case, without establishing that the same rule will catch every later failure.
968
+
969
+ The original and follow-up experiments use different models, workloads, reservations, and checkpoint schedules. Their memory growth cannot be combined into a universal allowance to add to a one-step peak. A short screen is useful for rejecting unsuitable settings; confidence in a chosen setting still requires evaluation over the operations and duration that matter for its intended use.
970
+
971
+ \section{Threats to validity}\label{threats-to-validity}
972
+
973
+ \subsection{Construct validity}\label{construct-validity}
974
+
975
+ Operational safety is defined by the chosen memory margin, not by a physical law. The sensitivity analysis shows how changing that convention changes the labels. External sampling may also miss peaks shorter than its interval, even when the calibration agrees for the sustained peaks observed in this study.
976
+
977
+ Stage attribution locates an observed failure; it does not identify the origin of every allocation that contributed to it. For example, memory retained from generation can contribute to a failure reported during weight synchronization. Similarly, agreement in whole-run outcomes under the instrumentation control does not establish that individual operations or stage timings are unaffected.
978
+
979
+ \subsection{Internal and statistical validity}\label{internal-and-statistical-validity}
980
+
981
+ The configuration levels change several controls together. Model comparisons also differ in architecture, parameter count, tokenizer, and sometimes the examples remaining after filtering. The batch/reservation experiment varies its two factors systematically but its original blocks are not controlled for allocation or execution order. The new same-GPU control addresses that concern for two selected cases, not the full factorial grid. Four-GPU measurements use a later experimental period rather than randomized GPU-count assignment.
982
+
983
+ Using the same GPUs and counterbalancing execution order reduces placement and order effects, but cannot eliminate changes in host caches, generated responses, or other conditions between runs. The original external-only startup timeouts and the new full-first noncompletion occur under different designs. Neither should be erased by successful later attempts or attributed to instrumentation without further evidence.
984
+
985
+ The workload set is small and selected rather than randomly sampled. Both code subsets come from the same corpus, and instrumentation base configurations share model--workload settings. Per-case reporting and corpus-weighting sensitivity make this dependence visible; repeated seeds and resampling cannot turn these cases into a representative population sample.
986
+
987
+ \subsection{External validity}\label{external-validity}
988
+
989
+ The tested models contain approximately 1.5--3.8 billion parameters. Results from this compact LoRA setup need not extend to larger models, full-parameter training, other accelerators, or different execution frameworks. The workloads also omit multimodal tasks, extended agent interactions, and very long contexts.
990
+
991
+ The longest runs cover 100 steps. They test repeated validation and checkpoint writing, but not restoration from checkpoints or memory behavior over a full training schedule. Those operations and longer durations need their own evaluation.
992
+
993
+ \subsection{Artifact validity}\label{artifact-validity}
994
+
995
+ The evidence package retains excluded attempts and earlier analyses, with records identifying the material used in this paper. File checksums can detect changes in those records, while reconstruction checks the analysis. Neither establishes that the original instrumentation captured every relevant event; that limitation remains separate from reproducibility.
996
+
997
+ \section{Reproducibility}\label{reproducibility}
998
+
999
+ The benchmark evidence is publicly available on Hugging Face \cite{kobza2026benchmark}, with a separate repository for reproduction software \cite{kobza2026reproduction}. Both can be accessed without authentication or approval. The software records the exact evidence version and checks its files before analysis.
1000
+
1001
+ The package contains the raw records described in the measurement protocol, experiment matrices, exclusion decisions, software environments, and scripts for the tables and figures. Reconstruction requires only a central processing unit (CPU), not a GPU training allocation. It starts from the raw evidence rather than using the published result tables as inputs, and checks run counts, seeds, stage records, and completed steps before comparing the regenerated results with the reference outputs.
1002
+
1003
+ The original evidence archive was downloaded and reconstructed by the authors in a separate environment where the source project and precomputed tables were unavailable to the analysis. All forty original result files matched after normalizing only file-location prefixes. This is an isolated reconstruction check, not an independent external replication. The revised release additionally contains the review-requested analyses and control records. Setup commands and detailed integrity reports accompany the reproduction code.
1004
+
1005
+ Rerunning GPU training is a separate task. The environment records provide the software and source versions, model and dataset identities, and launch settings needed for that purpose. The release excludes model weights and optimizer checkpoint payloads. Users obtain upstream models under their applicable terms; version and license records are supplied with the evidence.
1006
+
1007
+ \section{Conclusion}\label{conclusion}
1008
+
1009
+ RLVRAMBench provides a reproducible way to evaluate memory feasibility in colocated LoRA GRPO. Its measurements show why a successful run, an overall memory peak, and a label transferred from another setting can each give an incomplete answer. Retaining failures, measuring individual stages, and testing repeated and longer executions make those limitations visible.
1010
+
1011
+ Readers can reuse the protocol and linked records to compare configuration decisions, inspect failed attempts, and check whether a stage-level result survives a change in measurement or execution conditions. The controls demonstrate why that last check belongs in a memory benchmark alongside the reported peaks.
1012
+
1013
+ \section*{Acknowledgment}\label{acknowledgment}
1014
+ \addcontentsline{toc}{section}{Acknowledgment}
1015
+
1016
+ OpenAI Codex \cite{openai2026codex} was used to assist with coding and to refine and polish the manuscript text. The authors are responsible for the code, analyses, and content.
1017
+ \bibliographystyle{IEEEtran}
1018
+ \bibliography{references,ieee_access_additions}
1019
+ \phantomsection
1020
+ \begin{IEEEbiographynophoto}{Ond\v{r}ej Kobza}
1021
+ is affiliated with the Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague, Czech Republic. His work reported in this article concerns GPU-memory feasibility and reproducible benchmarking of large-language-model reinforcement learning.
1022
+ \end{IEEEbiographynophoto}
1023
+
1024
+ \phantomsection
1025
+ \begin{IEEEbiographynophoto}{Jan \v{S}ediv\'{y}}
1026
+ is affiliated with the Czech Institute of Informatics, Robotics and Cybernetics, Czech Technical University in Prague, Czech Republic. He supervises the research reported in this article.
1027
+ \end{IEEEbiographynophoto}
1028
+ \EOD
1029
+ \end{document}
paper/RLVRAMBench_supplement.pdf ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f012960eb466b9f01adf1dcd481f2c588a7198ad5dcf0b9692843064e387a099
3
+ size 220349
paper/RLVRAMBench_supplement.tex ADDED
@@ -0,0 +1,347 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ % Supplement to the directly edited IEEE Access submission.
2
+ \documentclass[10pt,a4paper]{article}
3
+ \usepackage[margin=22mm]{geometry}
4
+ \usepackage[T1]{fontenc}
5
+ \usepackage{lmodern,amsmath,booktabs,array,longtable}
6
+ \usepackage[hidelinks]{hyperref}
7
+ \usepackage{xurl}
8
+ \renewcommand{\thetable}{S\arabic{table}}
9
+ \renewcommand{\thesection}{S\arabic{section}}
10
+ \setlength{\parindent}{0pt}
11
+ \setlength{\parskip}{6pt}
12
+ \title{RLVRAMBench\\Supplementary Methods and Evidence}
13
+ \author{Ond\v{r}ej Kobza \and Jan \v{S}ediv\'{y}}
14
+ \date{}
15
+ \begin{document}
16
+ \maketitle
17
+
18
+ This supplement documents the finite-case comparisons and measurement checks
19
+ underlying the submission. It complements the article's configuration tables;
20
+ the public evidence and reproduction code contain the individual records.
21
+ Rounded values here are for reading. Analyses use the unrounded measurements.
22
+ Memory is reported in mebibytes (MiB, $2^{20}$ bytes).
23
+ Graphics processing units (GPUs) run the training and generation work;
24
+ CUDA is NVIDIA's execution platform. Other terminology follows the article.
25
+
26
+ \section{Attempt accounting and interpretation}
27
+
28
+ A planned slot identifies a configuration and seed, whereas an attempt is an
29
+ actual invocation of its runner. Queued jobs that never invoke the runner are
30
+ not process attempts. The original corpus contains 588 eligible slots and
31
+ 657 recorded attempts belonging to those slots. Eligibility is determined by
32
+ the documented artifact and failure rules, not by whether a measurement is
33
+ favorable.
34
+
35
+ The machine-readable inventory links each of the 69 excluded attempts to its
36
+ available evidence and original exclusion decision. Two are completed
37
+ processes from invalid paired allocations; one is a documented memory failure
38
+ with a corrupted trial record; 55 are other recorded failures; and eleven
39
+ have no trial record. These categories describe the available evidence.
40
+ They do not establish that every excluded failure has a known non-memory
41
+ cause.
42
+
43
+ The incomplete memory-failure record followed a terminal CUDA out-of-memory
44
+ error. A concurrent launcher edit prevented its normal trial record from
45
+ being written, although scheduler output retained the error. Its replacement
46
+ under identical scientific settings also failed. Restoring the known earlier
47
+ failure therefore leaves the repeated configuration label unchanged. It
48
+ would nevertheless be incorrect to describe the retained corpus as containing
49
+ every observed memory failure.
50
+
51
+ The original instrumentation control has two additional external-only
52
+ attempts that did not complete within the startup limit. Their cause is
53
+ unresolved and their full-instrumentation counterparts never started.
54
+ After separating two diagnosed setup attempts, the 36 eligible pairs and
55
+ these two incomplete pairs give a possible disagreement count between zero
56
+ and two across 38 allocation attempts. This range concerns unobserved
57
+ counterparts, not a measured treatment effect. No bootstrap of the eligible
58
+ zero-disagreement pairs resolves that missing information.
59
+
60
+ \section{Stage-specific effects of allocator instrumentation}
61
+
62
+ Both instrumentation conditions retain external device-memory sampling and
63
+ host-side stage markers. Full logging additionally waits for outstanding
64
+ GPU work before a traced worker operation, resets allocator peak counters,
65
+ and records allocator statistics when that operation returns. The wrapper
66
+ does not add a separate synchronization after the return.
67
+
68
+ Table S1 uses the common external traces, not the allocator's own memory
69
+ values. Each difference is full minus external-only within a same-GPU pair.
70
+ For a failed one-step run, a stage enters the contrast only when the trace
71
+ and diagnosed terminal stage show that execution advanced to a later
72
+ operation. The interrupted terminal stage remains a partial observation in
73
+ the raw record, not a completed-stage peak. A missing or interrupted stage
74
+ is not assigned a zero peak.
75
+
76
+ \begin{table}[htbp]
77
+ \centering
78
+ \caption{External-memory stage contrasts (MiB) in the original instrumentation
79
+ control. Entries are median [minimum, maximum] across three pairs.
80
+ Qwen, Phi, and Granite denote Qwen2.5-3B, Phi-4-mini, and Granite-3.3-2B.}
81
+ \small
82
+ \setlength{\tabcolsep}{4pt}
83
+ \begin{tabular}{@{}llrrr@{}}
84
+ \toprule
85
+ Model / workload & Level & Actor update & Generation & Weight synchronization \\
86
+ \midrule
87
+ Granite / GSM8K & c3 & $-114$ [$-424$, $-38$] & $-8$ [$-100$, 6] & 0 [0, 0] \\
88
+ Granite / GSM8K & c4 & $-120$ [$-334$, $-100$] & $-96$ [$-166$, $-6$] & 0 [0, 0] \\
89
+ Granite / MATH & c3 & $-60$ [$-198$, 56] & $-58$ [$-272$, 170] & 0 [0, 0] \\
90
+ Granite / MATH & c4 & 36 [$-216$, 140] & $-18$ [$-128$, 136] & 0 [0, 0] \\
91
+ Phi / GSM8K & c2 & 20 [$-1,800$, 1,456] & 22 [0, 28] & $-20$ [$-502$, 502] \\
92
+ Phi / GSM8K & c3 & 362 [$-320$, 550] & 24 [$-52$, 68] & 0 [$-20$, 608] \\
93
+ Phi / MATH & c2 & $-176$ [$-192$, 442] & $-46$ [$-402$, 76] & 0 [0, 0] \\
94
+ Phi / MATH & c3 & 398 [280, 798] & $-330$ [$-374$, 518] & 0 [0, 20] \\
95
+ Qwen / GSM8K & c3 & 116 [$-536$, 164] & 144 [86, 148] & 0 [0, 0] \\
96
+ Qwen / GSM8K & c4 & 400 [0, 404] & 0 [$-50$, 38] & Interrupted \\
97
+ Qwen / MATH & c3 & 60 [$-532$, 216] & $-22$ [$-42$, 106] & 0 [0, 0] \\
98
+ Qwen / MATH & c4 & Not observed & Not observed & Not observed \\
99
+ \bottomrule
100
+ \end{tabular}
101
+ \end{table}
102
+
103
+ Qwen/GSM8K c4 reaches actor update in all three pairs before failing.
104
+ Qwen/MATH c4 fails earlier. Thus, actor-stage contrasts use 33 pairs although
105
+ only thirty pairs complete in both conditions.
106
+
107
+ Sampled stage duration is calculated by assigning the interval between
108
+ successive timestamps to the earlier sample's marker. These durations are
109
+ available with the stage tables, but are not presented as precise operation
110
+ latencies: asynchronous work can cross marker boundaries.
111
+
112
+ \section{Factorial effects and the unit of comparison}
113
+
114
+ The original factorial study has 36 four-condition blocks: three models,
115
+ four workloads, and three seeds. Each block combines actor micro-batch sizes
116
+ 8 and 16 with generation-memory reservations 0.60 and 0.70. The main effect
117
+ of one factor averages its two matched differences over the other factor.
118
+ All four processes complete in every block.
119
+
120
+ Table S2 averages first within each workload, keeping model and seed
121
+ comparisons together. Both code workloads come from CodeContests; they
122
+ are not independent source corpora. Equal weighting of GSM8K, MATH, and
123
+ the mean of the two code effects reduces the actor-stage device-memory
124
+ effect from 3,980 to 3,574 MiB and the allocated-memory effect from 3,848 to
125
+ 3,450 MiB. The same weighting changes the reservation effect on the
126
+ whole-run peak from 3,998 to 4,006 MiB.
127
+
128
+ \begin{table}[htbp]
129
+ \centering
130
+ \caption{Per-workload factorial effects (MiB), nine blocks per workload.
131
+ All runs use full instrumentation. Device memory is sampled externally;
132
+ allocated memory is the distinct quantity recorded by the framework.}
133
+ \small
134
+ \begin{tabular}{@{}lrrrr@{}}
135
+ \toprule
136
+ & \multicolumn{3}{c}{Micro-batch 8 $\rightarrow$ 16} &
137
+ Reservation 0.60 $\rightarrow$ 0.70 \\
138
+ \cmidrule(lr){2-4}
139
+ Workload & Whole-run device & Actor device & Actor allocated & Whole-run device \\
140
+ \midrule
141
+ GSM8K & 0 & 2,073 & 2,000 & 4,017 \\
142
+ MATH & $-13$ & 3,448 & 3,311 & 4,030 \\
143
+ Standard code & $-14$ & 4,132 & 4,006 & 4,064 \\
144
+ Longer-prompt code & $-95$ & 6,268 & 6,073 & 3,881 \\
145
+ \bottomrule
146
+ \end{tabular}
147
+ \end{table}
148
+
149
+ The analytical blocks should not be confused with controlled allocations.
150
+ The original 144 processes ran in separate allocations on twenty nodes,
151
+ with unrandomized placement and execution order. Their environment records
152
+ include source revisions and working-copy status; some checkouts were
153
+ modified. The new batch-by-logging experiment addresses allocation and
154
+ instrumentation sensitivity in two selected cases, rather than retrospectively
155
+ turning the original study into a same-GPU experiment.
156
+
157
+ \section{Realized work in longer runs}
158
+
159
+ The article reports realized lengths for every model--workload group in the
160
+ boundary experiment. Table S3 extends this audit to completed 100-step runs.
161
+ The original two cases have two regimes with three seeds each; the follow-up
162
+ has one regime and three seeds per case. Means weight completed processes
163
+ equally. Every completed process has a length summary and a positive
164
+ gradient-norm measurement at every requested step. These are checks of
165
+ training progress, not evidence of improved task accuracy.
166
+
167
+ \begin{table}[htbp]
168
+ \centering
169
+ \caption{Actual sequence lengths in the 100-step studies. Lengths are tokens.
170
+ ``Cap batches'' counts batches containing at least one response reaching its
171
+ configured cap, not the percentage of responses at the cap.}
172
+ \small
173
+ \begin{tabular}{@{}llrrrr@{}}
174
+ \toprule
175
+ Study / model & Workload & Runs & Prompt mean / max. & Response mean / max. & Cap batches \\
176
+ \midrule
177
+ Original / Qwen2.5-1.5B & MATH & 6 & 120 / 1,698 & 536 / 1,024 & 600 / 600 \\
178
+ Original / Qwen2.5-3B & GSM8K & 6 & 104 / 256 & 267 / 1,024 & 50 / 600 \\
179
+ Follow-up / Granite & GSM8K & 3 & 141 / 296 & 265 / 2,048 & 107 / 300 \\
180
+ Follow-up / Granite & MATH & 3 & 145 / 512 & 405 / 512 & 300 / 300 \\
181
+ Follow-up / Phi & GSM8K & 3 & 73 / 200 & 216 / 2,048 & 45 / 300 \\
182
+ Follow-up / Phi & MATH & 3 & 84 / 498 & 408 / 512 & 300 / 300 \\
183
+ \bottomrule
184
+ \end{tabular}
185
+ \end{table}
186
+
187
+ The trainer's metric named \texttt{response\_length/clip\_ratio} compares
188
+ response length with the width of that batch's response tensor. That width
189
+ can be smaller than the configured generation cap. It is therefore not used
190
+ as an estimate of the fraction of responses truncated by the configured cap.
191
+ The audit instead compares the logged batch maximum with the configured
192
+ limit, giving the counts in Table S3.
193
+
194
+ Training data use fixed order and incomplete final batches are dropped.
195
+ For filtered dataset size $N$, global batch size $B$, and $T$ completed
196
+ steps, assuming $N\geq B>0$, the number of distinct examples implied by this schedule is
197
+ \[
198
+ B\min\left(T,\left\lfloor N/B\right\rfloor\right),
199
+ \]
200
+ with $\lfloor(T-1)/\lfloor N/B\rfloor\rfloor$ restarts of the data iterator.
201
+ These are schedule-derived exposure counts, not an audit of recorded
202
+ per-example identities.
203
+
204
+ Each original 100-step run processes 7,424 distinct examples under this
205
+ calculation and restarts the fixed order once. Its filtered dataset has
206
+ 7,473 GSM8K or 7,500 MATH examples. The follow-up uses 3,200 distinct GSM8K
207
+ examples or 6,400 MATH examples without a restart. Its filtered MATH sets
208
+ have 7,395 examples for Granite and 7,434 for Phi, whereas both GSM8K sets
209
+ have 7,473. Consequently, nominally matching a workload does not guarantee
210
+ identical token sequences or example identities across models.
211
+
212
+ \section{Original temporal effects by case}
213
+
214
+ Later growth compares each 100-step run with its own first-step peak.
215
+ The validation/checkpoint contrast subtracts the maximum of all other
216
+ stages from the larger validation or checkpoint peak. A negative value
217
+ means these explicitly tested operations do not set the overall maximum.
218
+ Table S4 reports observed variation across the six runs in each case,
219
+ without a two-case population confidence interval.
220
+
221
+ \begin{table}[htbp]
222
+ \centering
223
+ \caption{Original 100-step effects by case (MiB). Brackets are observed run
224
+ ranges, not confidence intervals.}
225
+ \small
226
+ \begin{tabular}{@{}lrr@{}}
227
+ \toprule
228
+ Case & Later growth: mean [range] &
229
+ Validation/checkpoint contrast: mean [range] \\
230
+ \midrule
231
+ Qwen2.5-1.5B / MATH & 2,083 [1,480 to 2,448] & $-$2,741 [$-$3,440 to $-$2,070] \\
232
+ Qwen2.5-3B / GSM8K & 1,767 [1,464 to 2,168] & $-$1,963 [$-$2,508 to $-$1,722] \\
233
+ \bottomrule
234
+ \end{tabular}
235
+ \end{table}
236
+
237
+ \section{Targeted batch-by-logging control}
238
+
239
+ The initial design contains six four-condition allocations: three seeds in
240
+ Phi/GSM8K and three in Qwen2.5-3B/longer-prompt code. Conditions E8 and E16
241
+ use external-only measurement with actor micro-batches 8 and 16; F8 and F16
242
+ use full allocator logging at those batch sizes. All other scientific
243
+ settings are fixed within a case, with reservation 0.60. Every condition
244
+ starts in a fresh process after the same two GPUs return to the prescribed
245
+ idle-memory baseline.
246
+
247
+ Four initial allocations completed all conditions. The Phi seed122
248
+ allocation stopped in F16, its first condition, after an accelerator
249
+ argument error during weight-transfer buffer reconstruction and a
250
+ 1,200-second timeout. The Phi seed123 allocation stopped in F8, also first,
251
+ at the same timeout with unresolved cause. These are observed
252
+ noncompletions; neither is established as an out-of-memory failure. The
253
+ other six conditions in those allocations never started.
254
+
255
+ A dated amendment was recorded after those outcomes and before the
256
+ additional runs. It allowed at most one new allocation for each incomplete
257
+ block without a diagnosed memory failure, with the same source, settings,
258
+ seed, order, and timeout. Both additional allocations completed. The
259
+ experiment therefore records 26 attempted processes: sixteen completions
260
+ and two noncompletions initially, followed by eight completions in the
261
+ bounded extension. All 24 completed processes stay within margin.
262
+
263
+ Table S5 keeps every continuous contrast within one allocation. It does
264
+ not splice an initial condition into a later block. The two additional
265
+ allocations supply conditional evidence about memory use; they do not
266
+ replace the initial execution outcomes or increase the number of seeds.
267
+
268
+ \begin{table}[htbp]
269
+ \centering
270
+ \caption{All usable control blocks. Memory effects are MiB. An asterisk
271
+ identifies the single additional allocation after initial noncompletion.
272
+ The last column subtracts the external-only batch effect from the full
273
+ logging batch effect.}
274
+ \small
275
+ \setlength{\tabcolsep}{4pt}
276
+ \begin{tabular}{@{}lr lrrr@{}}
277
+ \toprule
278
+ Case & Seed & Condition order & E16$-$E8 & F16$-$F8 & Difference \\
279
+ \midrule
280
+ Phi / GSM8K & 121 & E8, F16, F8, E16 & 2,538 & 3,308 & 770 \\
281
+ Phi / GSM8K & 122* & F16, E8, E16, F8 & 652 & 680 & 28 \\
282
+ Phi / GSM8K & 123* & F8, E16, E8, F16 & 2,708 & 2,874 & 166 \\
283
+ Qwen / longer-prompt code & 121 & E16, F8, F16, E8 & 6,068 & 7,120 & 1,052 \\
284
+ Qwen / longer-prompt code & 122 & E8, E16, F16, F8 & 5,922 & 6,976 & 1,054 \\
285
+ Qwen / longer-prompt code & 123 & F16, F8, E8, E16 & 7,300 & 7,080 & $-220$ \\
286
+ \bottomrule
287
+ \end{tabular}
288
+ \end{table}
289
+
290
+ Actor-stage batch effects are positive under both measurement conditions
291
+ in every usable block. The external-only means are 1,966 MiB for Phi and
292
+ 6,430 MiB for Qwen. Full logging gives 2,287 and 7,059 MiB, respectively.
293
+ Three block differences exceed the pre-execution 512 MiB diagnostic in
294
+ absolute value; that reference is not an equivalence margin established
295
+ by these two cases. All whole-run batch effects are between $-160$ and
296
+ 132 MiB, so the positive actor-stage effect is not an artifact of reporting
297
+ the overall maximum.
298
+
299
+ Table S6 shows why these data should not be interpreted as tracing overhead
300
+ on an identical tensor workload. Prompts have the same mean length in
301
+ every condition of a case: 78.1 tokens for Phi/GSM8K and 720.2 for
302
+ Qwen/longer-prompt code. Generated responses differ despite the matched
303
+ seed. In particular, both batch-8 Phi seed122 conditions contain a much
304
+ longer response than their batch-16 counterparts. Every completed condition
305
+ records one training step and a positive gradient norm.
306
+
307
+ \begin{table}[htbp]
308
+ \centering
309
+ \caption{Realized response lengths for the blocks in Table S5. Entries are
310
+ mean / maximum tokens, rounded to whole tokens.}
311
+ \small
312
+ \begin{tabular}{@{}lrrrrr@{}}
313
+ \toprule
314
+ Case & Seed & E8 & E16 & F8 & F16 \\
315
+ \midrule
316
+ Phi / GSM8K & 121 & 222 / 505 & 228 / 414 & 213 / 403 & 231 / 625 \\
317
+ Phi / GSM8K & 122* & 267 / 2,048 & 219 / 455 & 275 / 2,048 & 212 / 442 \\
318
+ Phi / GSM8K & 123* & 209 / 497 & 230 / 456 & 217 / 363 & 220 / 503 \\
319
+ Qwen / longer-prompt code & 121 & 357 / 1,024 & 326 / 1,024 & 346 / 728 & 351 / 1,024 \\
320
+ Qwen / longer-prompt code & 122 & 373 / 1,024 & 333 / 739 & 330 / 1,024 & 330 / 990 \\
321
+ Qwen / longer-prompt code & 123 & 334 / 760 & 385 / 1,024 & 355 / 1,024 & 364 / 1,024 \\
322
+ \bottomrule
323
+ \end{tabular}
324
+ \end{table}
325
+
326
+ The initial noncompletions both occurred in full-logging conditions in first
327
+ position. Logging, order, seed, and allocation cannot be disentangled as
328
+ causes from these observations. Successful additional attempts do not
329
+ establish startup equivalence. The supported result is narrower: increasing
330
+ the training micro-batch raises the actor-stage peak under external-only
331
+ measurement in every observed block, while its magnitude varies with
332
+ execution and generated work.
333
+
334
+ \section{Interpretation of reconstruction}
335
+
336
+ The original 588 completed or memory-failing records contain 504 completed
337
+ processes with 3,465 logged training steps. All of those steps have positive
338
+ gradient-norm measurements. Supporting historical screens and abandoned-scope
339
+ observations are separate from that count.
340
+
341
+ Reconstruction checks recorded evidence, not the completeness of the original
342
+ instrumentation. In particular, external sampling may miss a short transient,
343
+ an error's reported stage need not be where contributing memory was allocated,
344
+ and immutable archived files do not imply that every historical source
345
+ checkout was immutable during execution. The attempt inventory, environment
346
+ records, and new controls expose these distinctions rather than remove them.
347
+ \end{document}
paper/figures/review_instrumentation_batch.pdf ADDED
Binary file (15.9 kB). View file
 
paper/figures/review_instrumentation_batch.png ADDED

Git LFS Details

  • SHA256: 23b742462ff86cf810dda6413d35e5d4678bc9f6c7412dbf893f6a70e2732771
  • Pointer size: 131 Bytes
  • Size of remote file: 116 kB
paper/figures/standard_grpo_boundary.pdf ADDED
Binary file (15.8 kB). View file
 
paper/figures/standard_grpo_boundary.png ADDED

Git LFS Details

  • SHA256: 7b2d35dd05bad3848d8dd60fa8b2d6e07f3f42d131767def46401510f23eeeee
  • Pointer size: 131 Bytes
  • Size of remote file: 185 kB
paper/figures/standard_grpo_mechanisms.pdf ADDED
Binary file (17.6 kB). View file
 
paper/figures/standard_grpo_mechanisms.png ADDED

Git LFS Details

  • SHA256: f3ee3f83a769c300e9527381ed90d8bb605debed14333ac8a6e024322a59f66a
  • Pointer size: 131 Bytes
  • Size of remote file: 189 kB
paper/figures/standard_grpo_temporal.pdf ADDED
Binary file (17 kB). View file
 
paper/figures/standard_grpo_temporal.png ADDED

Git LFS Details

  • SHA256: 5760e69dcf7d59885376fe734c8f642ba448a29074805d6924c175f85a1b5d3f
  • Pointer size: 130 Bytes
  • Size of remote file: 98.4 kB
release.json CHANGED
@@ -1,15 +1,15 @@
1
  {
2
- "release": "ieee-access-20260914-2897476",
3
- "source_commit": "289747694f1a71052acc7977e742bfc0a45286a2",
4
  "huggingface_repository": "kobzaond/RLVRAMBench",
5
  "code_repository": "kobzaond/RLVRAMBench-reproduce",
6
  "archive": {
7
- "path": "releases/RLVRAMBench-review-artifact-ieee-access-20260914-2897476.tar.zst",
8
- "sha256": "4a735cdc33b709d330b9894cb3b923542da2c29c29d366eccdae19c781096cfd",
9
- "size_bytes": 171120507
10
  },
11
- "manifest_entries": 19226,
12
- "archive_members": 19228,
13
  "expected_derived_files": 40,
14
  "principal_processes": 588,
15
  "supporting_source_processes": 16,
@@ -18,5 +18,12 @@
18
  "visibility_at_initial_deposit": "private",
19
  "visibility": "public",
20
  "public_release_date": "2026-09-14",
21
- "original_material_license": null
 
 
 
 
 
 
 
22
  }
 
1
  {
2
+ "release": "critical-revision-20260914-328c1c3",
3
+ "source_commit": "328c1c33a8b4353429be8e892ee6f8b3eb57c1c2",
4
  "huggingface_repository": "kobzaond/RLVRAMBench",
5
  "code_repository": "kobzaond/RLVRAMBench-reproduce",
6
  "archive": {
7
+ "path": "releases/RLVRAMBench-review-artifact-critical-revision-20260914-328c1c3.tar.zst",
8
+ "sha256": "81f9703da21b1fda72627960e477acdca5d196995787272d093788d46e6b9e4d",
9
+ "size_bytes": 174064706
10
  },
11
+ "manifest_entries": 19509,
12
+ "archive_members": 19511,
13
  "expected_derived_files": 40,
14
  "principal_processes": 588,
15
  "supporting_source_processes": 16,
 
18
  "visibility_at_initial_deposit": "private",
19
  "visibility": "public",
20
  "public_release_date": "2026-09-14",
21
+ "original_material_license": null,
22
+ "expected_review_derived_files": 27,
23
+ "review_control_completed_processes": 24,
24
+ "review_control_attempted_processes": 26,
25
+ "review_control_usable_blocks": 6,
26
+ "original_recorded_attempts": 657,
27
+ "prior_immutable_release_preserved": "ieee-access-20260914-2897476",
28
+ "manuscript_editing_authority": "direct_tex"
29
  }
releases/RLVRAMBench-release-critical-revision-20260914-328c1c3.json ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "archives": [
3
+ {
4
+ "path": "/mnt/proj3/open-35-44/verl/releases/rlvrambench/RLVRAMBench-review-artifact-critical-revision-20260914-328c1c3.tar.zst",
5
+ "sha256": "81f9703da21b1fda72627960e477acdca5d196995787272d093788d46e6b9e4d",
6
+ "size_bytes": 174064706
7
+ },
8
+ {
9
+ "path": "/mnt/proj3/open-35-44/verl/releases/rlvrambench/RLVRAMBench-submission-scientific-package-critical-revision-20260914-328c1c3.zip",
10
+ "sha256": "89b5c3068d9e0c9f6fc2cc1faebfa25b628896e3113696688bcbe254c1e16ad2",
11
+ "size_bytes": 1783963
12
+ }
13
+ ],
14
+ "author_supplied_upload_blockers": [
15
+ "Final author order, affiliations, postal addresses, and corresponding author details",
16
+ "CRediT roles, funding, competing interests, and author approval",
17
+ "Review artifact URL, permanent DOI, and repository-level license",
18
+ "Final referee suggestions and submission-system data statement"
19
+ ],
20
+ "manifest_entries": 19509,
21
+ "review_members": 19511,
22
+ "submission_members": 3,
23
+ "tag": "critical-revision-20260914-328c1c3"
24
+ }
releases/RLVRAMBench-review-artifact-critical-revision-20260914-328c1c3.tar.zst ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:81f9703da21b1fda72627960e477acdca5d196995787272d093788d46e6b9e4d
3
+ size 174064706
releases/RLVRAMBench-submission-scientific-package-critical-revision-20260914-328c1c3.zip ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:89b5c3068d9e0c9f6fc2cc1faebfa25b628896e3113696688bcbe254c1e16ad2
3
+ size 1783963
tables/review_revision/attempt-flow.csv ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ study,planned_slots,recorded_allocation_attempts,recorded_process_attempts,retained,retained_completed,retained_memory_failure,excluded_process_attempts,excluded_completed,excluded_memory_failure,excluded_other_failure,excluded_no_trial_record
2
+ boundary,216,251,251,216,156,60,35,0,1,33,1
3
+ state_release,36,21,41,36,27,9,5,2,0,2,1
4
+ factorial,144,161,161,144,144,0,17,0,0,15,2
5
+ original_40_step,18,18,18,18,15,3,0,0,0,0,0
6
+ original_100_step,12,14,14,12,12,0,2,0,0,0,2
7
+ sampling_calibration,12,12,12,12,12,0,0,0,0,0,0
8
+ instrumentation,72,40,76,72,60,12,4,0,0,1,3
9
+ followup_100_step,24,15,27,24,24,0,3,0,0,2,1
10
+ gpu_count,54,57,57,54,54,0,3,0,0,2,1
tables/review_revision/attempt-workload-summary.json ADDED
@@ -0,0 +1,190 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "principal_included_slots": 588,
3
+ "recorded_process_attempts": 657,
4
+ "excluded_attempts": 69,
5
+ "excluded_reasons": {
6
+ "launcher-mutation race while the shared instrumented launcher was amended after a terminal vLLM CUDA OOM; the attempt retained scheduler stdout, NVML telemetry, phase telemetry, environment metadata, and allocator stubs, but the mixed shell read truncated the designated training log and prevented trial JSON creation": 1,
7
+ "vLLM worker TCPStore port collision during distributed initialization (EADDRINUSE); the failure occurred before the requested training step and produced no scientific memory outcome": 1,
8
+ "granite_live_launcher_replacement_failure": 8,
9
+ "granite_vllm_replica_port_collision": 24,
10
+ "granite_startup_deadlock": 1,
11
+ "same_node_pair_infrastructure_failure": 2,
12
+ "same_node_pair_port_collision": 2,
13
+ "same_node_period1_startup_deadlock": 1,
14
+ "factorial_vllm_replica_port_collision": 8,
15
+ "factorial_live_launcher_replacement_failure": 6,
16
+ "factorial_startup_deadlock": 2,
17
+ "factorial_ray_startup_segfault": 1,
18
+ "temporal_cuda_ipc_weight_sync_failure": 1,
19
+ "temporal_weight_sync_no_progress": 1,
20
+ "revision_snapshot_missing_submodule": 3,
21
+ "revision_snapshot_missing_hydra_data_config": 3,
22
+ "revision_startup_stall_timeout": 2,
23
+ "revision_native_phi_checkpoint_metadata": 1,
24
+ "revision_four_gpu_port_span": 1
25
+ },
26
+ "flow": [
27
+ {
28
+ "study": "boundary",
29
+ "planned_slots": 216,
30
+ "recorded_allocation_attempts": 251,
31
+ "recorded_process_attempts": 251,
32
+ "retained": 216,
33
+ "retained_completed": 156,
34
+ "retained_memory_failure": 60,
35
+ "excluded_process_attempts": 35,
36
+ "excluded_completed": 0,
37
+ "excluded_memory_failure": 1,
38
+ "excluded_other_failure": 33,
39
+ "excluded_no_trial_record": 1
40
+ },
41
+ {
42
+ "study": "state_release",
43
+ "planned_slots": 36,
44
+ "recorded_allocation_attempts": 21,
45
+ "recorded_process_attempts": 41,
46
+ "retained": 36,
47
+ "retained_completed": 27,
48
+ "retained_memory_failure": 9,
49
+ "excluded_process_attempts": 5,
50
+ "excluded_completed": 2,
51
+ "excluded_memory_failure": 0,
52
+ "excluded_other_failure": 2,
53
+ "excluded_no_trial_record": 1
54
+ },
55
+ {
56
+ "study": "factorial",
57
+ "planned_slots": 144,
58
+ "recorded_allocation_attempts": 161,
59
+ "recorded_process_attempts": 161,
60
+ "retained": 144,
61
+ "retained_completed": 144,
62
+ "retained_memory_failure": 0,
63
+ "excluded_process_attempts": 17,
64
+ "excluded_completed": 0,
65
+ "excluded_memory_failure": 0,
66
+ "excluded_other_failure": 15,
67
+ "excluded_no_trial_record": 2
68
+ },
69
+ {
70
+ "study": "original_40_step",
71
+ "planned_slots": 18,
72
+ "recorded_allocation_attempts": 18,
73
+ "recorded_process_attempts": 18,
74
+ "retained": 18,
75
+ "retained_completed": 15,
76
+ "retained_memory_failure": 3,
77
+ "excluded_process_attempts": 0,
78
+ "excluded_completed": 0,
79
+ "excluded_memory_failure": 0,
80
+ "excluded_other_failure": 0,
81
+ "excluded_no_trial_record": 0
82
+ },
83
+ {
84
+ "study": "original_100_step",
85
+ "planned_slots": 12,
86
+ "recorded_allocation_attempts": 14,
87
+ "recorded_process_attempts": 14,
88
+ "retained": 12,
89
+ "retained_completed": 12,
90
+ "retained_memory_failure": 0,
91
+ "excluded_process_attempts": 2,
92
+ "excluded_completed": 0,
93
+ "excluded_memory_failure": 0,
94
+ "excluded_other_failure": 0,
95
+ "excluded_no_trial_record": 2
96
+ },
97
+ {
98
+ "study": "sampling_calibration",
99
+ "planned_slots": 12,
100
+ "recorded_allocation_attempts": 12,
101
+ "recorded_process_attempts": 12,
102
+ "retained": 12,
103
+ "retained_completed": 12,
104
+ "retained_memory_failure": 0,
105
+ "excluded_process_attempts": 0,
106
+ "excluded_completed": 0,
107
+ "excluded_memory_failure": 0,
108
+ "excluded_other_failure": 0,
109
+ "excluded_no_trial_record": 0
110
+ },
111
+ {
112
+ "study": "instrumentation",
113
+ "planned_slots": 72,
114
+ "recorded_allocation_attempts": 40,
115
+ "recorded_process_attempts": 76,
116
+ "retained": 72,
117
+ "retained_completed": 60,
118
+ "retained_memory_failure": 12,
119
+ "excluded_process_attempts": 4,
120
+ "excluded_completed": 0,
121
+ "excluded_memory_failure": 0,
122
+ "excluded_other_failure": 1,
123
+ "excluded_no_trial_record": 3
124
+ },
125
+ {
126
+ "study": "followup_100_step",
127
+ "planned_slots": 24,
128
+ "recorded_allocation_attempts": 15,
129
+ "recorded_process_attempts": 27,
130
+ "retained": 24,
131
+ "retained_completed": 24,
132
+ "retained_memory_failure": 0,
133
+ "excluded_process_attempts": 3,
134
+ "excluded_completed": 0,
135
+ "excluded_memory_failure": 0,
136
+ "excluded_other_failure": 2,
137
+ "excluded_no_trial_record": 1
138
+ },
139
+ {
140
+ "study": "gpu_count",
141
+ "planned_slots": 54,
142
+ "recorded_allocation_attempts": 57,
143
+ "recorded_process_attempts": 57,
144
+ "retained": 54,
145
+ "retained_completed": 54,
146
+ "retained_memory_failure": 0,
147
+ "excluded_process_attempts": 3,
148
+ "excluded_completed": 0,
149
+ "excluded_memory_failure": 0,
150
+ "excluded_other_failure": 2,
151
+ "excluded_no_trial_record": 1
152
+ }
153
+ ],
154
+ "instrumentation_attempts_by_condition": {
155
+ "external": {
156
+ "recorded_attempts": 40,
157
+ "retained": 36,
158
+ "unresolved_timeout_attempts": 2,
159
+ "other_excluded_attempts": 2
160
+ },
161
+ "full": {
162
+ "recorded_attempts": 36,
163
+ "retained": 36,
164
+ "unresolved_timeout_attempts": 0,
165
+ "other_excluded_attempts": 0
166
+ }
167
+ },
168
+ "unknown_timeout_pair_sensitivity": {
169
+ "eligible_pairs": 36,
170
+ "eligible_discordances": 0,
171
+ "additional_unknown_external_first_pairs": 2,
172
+ "observed_counterpart_for_unknown_pairs": false,
173
+ "possible_discordances_over_38_pair_attempts": [
174
+ 0,
175
+ 2
176
+ ],
177
+ "interpretation": "Missing-data range, not a causal effect or confidence interval"
178
+ },
179
+ "length_metric_warning": "VERL clip_ratio compares length with the batch tensor width, not necessarily the configured generation cap. Report the number of logged batches whose maximum reaches the configured cap, not a fabricated per-response cap fraction.",
180
+ "distinct_examples_warning": "Derived from recorded fixed data order, filtered dataset size, batch size, drop_last=True, and logged steps; not a per-example ID audit.",
181
+ "attempt_inventory_scope": "Recorded runner attempts belonging to the 588 planned reported slots. Supporting historical screens and abandoned-scope experiments are separate. Queued allocations never reaching a runner are not processes.",
182
+ "outputs": {
183
+ "recorded-attempts": 657,
184
+ "attempt-flow": 9,
185
+ "realized-workload-trials": 588,
186
+ "realized-workload-cases": 52,
187
+ "state-release-design": 6,
188
+ "long-run-exposure": 6
189
+ }
190
+ }
tables/review_revision/control-allocation-attempts.csv ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ manifest,pair_id,job_id,status,recorded_periods,error
2
+ output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s121/pair-5048502.json,phi4_mini-gsm8k-s121,5048502,complete,4,
3
+ output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s122/pair-5048503.json,phi4_mini-gsm8k-s122,5048503,infrastructure_invalid,1,RuntimeError: no trial outcome for review-control-phi4_mini-gsm8k-s122-full_b16
4
+ output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s122/pair-5048577.json,phi4_mini-gsm8k-s122,5048577,complete,4,
5
+ output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s123/pair-5048508.json,phi4_mini-gsm8k-s123,5048508,infrastructure_invalid,1,RuntimeError: no trial outcome for review-control-phi4_mini-gsm8k-s123-full_b8
6
+ output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s123/pair-5048576.json,phi4_mini-gsm8k-s123,5048576,complete,4,
7
+ output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s121/pair-5048525.json,qwen25_3b-code_heavy_tail-s121,5048525,complete,4,
8
+ output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s122/pair-5048560.json,qwen25_3b-code_heavy_tail-s122,5048560,complete,4,
9
+ output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s123/pair-5048501.json,qwen25_3b-code_heavy_tail-s123,5048501,complete,4,
tables/review_revision/control-blocks.csv ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ pair_id,model_family,dataset,training_seed,job_id,order,initial_allocation,completed,within_margin,actor_update_external_batch_effect_mib,actor_update_full_batch_effect_mib,actor_update_instrumentation_interaction_mib,rollout_external_batch_effect_mib,rollout_full_batch_effect_mib,rollout_instrumentation_interaction_mib,weight_sync_external_batch_effect_mib,weight_sync_full_batch_effect_mib,weight_sync_instrumentation_interaction_mib,whole_run_external_batch_effect_mib,whole_run_full_batch_effect_mib,whole_run_instrumentation_interaction_mib
2
+ phi4_mini-gsm8k-s121,phi4_mini,gsm8k,121,5048502,external_b8;full_b16;full_b8;external_b16,1,4,4,2538.0,3308.0,770.0,2.0,22.0,20.0,0.0,0.0,0.0,-20.0,0.0,20.0
3
+ phi4_mini-gsm8k-s122,phi4_mini,gsm8k,122,5048577,full_b16;external_b8;external_b16;full_b8,0,4,4,652.0,680.0,28.0,-8.0,-6.0,2.0,0.0,-1004.0,-1004.0,0.0,0.0,0.0
4
+ phi4_mini-gsm8k-s123,phi4_mini,gsm8k,123,5048576,full_b8;external_b16;external_b8;full_b16,0,4,4,2708.0,2874.0,166.0,-26.0,-4.0,22.0,20.0,0.0,-20.0,20.0,0.0,-20.0
5
+ qwen25_3b-code_heavy_tail-s121,qwen25_3b,code_heavy_tail,121,5048525,external_b16;full_b8;full_b16;external_b8,1,4,4,6068.0,7120.0,1052.0,100.0,0.0,-100.0,-1446.0,0.0,1446.0,100.0,0.0,-100.0
6
+ qwen25_3b-code_heavy_tail-s122,qwen25_3b,code_heavy_tail,122,5048560,external_b8;external_b16;full_b16;full_b8,1,4,4,5922.0,6976.0,1054.0,-160.0,132.0,292.0,0.0,964.0,964.0,-160.0,132.0,292.0
7
+ qwen25_3b-code_heavy_tail-s123,qwen25_3b,code_heavy_tail,123,5048501,full_b16;full_b8;external_b8;external_b16,1,4,4,7300.0,7080.0,-220.0,-98.0,0.0,98.0,0.0,0.0,0.0,-98.0,0.0,98.0
tables/review_revision/control-cases.csv ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ model_family,dataset,blocks,actor_update_external_batch_effect_mib_observed_blocks,actor_update_external_batch_effect_mib_mean,actor_update_external_batch_effect_mib_min,actor_update_external_batch_effect_mib_max,actor_update_full_batch_effect_mib_observed_blocks,actor_update_full_batch_effect_mib_mean,actor_update_full_batch_effect_mib_min,actor_update_full_batch_effect_mib_max,actor_update_instrumentation_interaction_mib_observed_blocks,actor_update_instrumentation_interaction_mib_mean,actor_update_instrumentation_interaction_mib_min,actor_update_instrumentation_interaction_mib_max,rollout_external_batch_effect_mib_observed_blocks,rollout_external_batch_effect_mib_mean,rollout_external_batch_effect_mib_min,rollout_external_batch_effect_mib_max,rollout_full_batch_effect_mib_observed_blocks,rollout_full_batch_effect_mib_mean,rollout_full_batch_effect_mib_min,rollout_full_batch_effect_mib_max,rollout_instrumentation_interaction_mib_observed_blocks,rollout_instrumentation_interaction_mib_mean,rollout_instrumentation_interaction_mib_min,rollout_instrumentation_interaction_mib_max,weight_sync_external_batch_effect_mib_observed_blocks,weight_sync_external_batch_effect_mib_mean,weight_sync_external_batch_effect_mib_min,weight_sync_external_batch_effect_mib_max,weight_sync_full_batch_effect_mib_observed_blocks,weight_sync_full_batch_effect_mib_mean,weight_sync_full_batch_effect_mib_min,weight_sync_full_batch_effect_mib_max,weight_sync_instrumentation_interaction_mib_observed_blocks,weight_sync_instrumentation_interaction_mib_mean,weight_sync_instrumentation_interaction_mib_min,weight_sync_instrumentation_interaction_mib_max,whole_run_external_batch_effect_mib_observed_blocks,whole_run_external_batch_effect_mib_mean,whole_run_external_batch_effect_mib_min,whole_run_external_batch_effect_mib_max,whole_run_full_batch_effect_mib_observed_blocks,whole_run_full_batch_effect_mib_mean,whole_run_full_batch_effect_mib_min,whole_run_full_batch_effect_mib_max,whole_run_instrumentation_interaction_mib_observed_blocks,whole_run_instrumentation_interaction_mib_mean,whole_run_instrumentation_interaction_mib_min,whole_run_instrumentation_interaction_mib_max
2
+ phi4_mini,gsm8k,3,3,1966.0,652.0,2708.0,3,2287.3333333333335,680.0,3308.0,3,321.3333333333333,28.0,770.0,3,-10.666666666666666,-26.0,2.0,3,4.0,-6.0,22.0,3,14.666666666666666,2.0,22.0,3,6.666666666666667,0.0,20.0,3,-334.6666666666667,-1004.0,0.0,3,-341.3333333333333,-1004.0,0.0,3,0.0,-20.0,20.0,3,0.0,0.0,0.0,3,0.0,-20.0,20.0
3
+ qwen25_3b,code_heavy_tail,3,3,6430.0,5922.0,7300.0,3,7058.666666666667,6976.0,7120.0,3,628.6666666666666,-220.0,1054.0,3,-52.666666666666664,-160.0,100.0,3,44.0,0.0,132.0,3,96.66666666666667,-100.0,292.0,3,-482.0,-1446.0,0.0,3,321.3333333333333,0.0,964.0,3,803.3333333333334,0.0,1446.0,3,-52.666666666666664,-160.0,100.0,3,44.0,0.0,132.0,3,96.66666666666667,-100.0,292.0
tables/review_revision/control-process-attempts.csv ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_id,pair_id,condition,period,model_family,dataset,training_seed,job_id,initial_allocation,attempted,outcome,observed_peak_mib,actor_update_peak_mib,terminal_sampled_phase
2
+ review-control-phi4_mini-gsm8k-s121-external_b8,phi4_mini-gsm8k-s121,external_b8,1,phi4_mini,gsm8k,121,5048502,1,1,completed,35538.0,16824.0,idle
3
+ review-control-phi4_mini-gsm8k-s121-full_b16,phi4_mini-gsm8k-s121,full_b16,2,phi4_mini,gsm8k,121,5048502,1,1,completed,35518.0,19760.0,idle
4
+ review-control-phi4_mini-gsm8k-s121-full_b8,phi4_mini-gsm8k-s121,full_b8,3,phi4_mini,gsm8k,121,5048502,1,1,completed,35518.0,16452.0,idle
5
+ review-control-phi4_mini-gsm8k-s121-external_b16,phi4_mini-gsm8k-s121,external_b16,4,phi4_mini,gsm8k,121,5048502,1,1,completed,35518.0,19362.0,idle
6
+ review-control-phi4_mini-gsm8k-s122-full_b16,phi4_mini-gsm8k-s122,full_b16,1,phi4_mini,gsm8k,122,5048503,1,1,accelerator_argument_error_then_noncompletion,35398.0,nan,initializing
7
+ review-control-phi4_mini-gsm8k-s122-external_b8,phi4_mini-gsm8k-s122,external_b8,2,phi4_mini,gsm8k,122,5048503,1,0,not_attempted,nan,nan,
8
+ review-control-phi4_mini-gsm8k-s122-external_b16,phi4_mini-gsm8k-s122,external_b16,3,phi4_mini,gsm8k,122,5048503,1,0,not_attempted,nan,nan,
9
+ review-control-phi4_mini-gsm8k-s122-full_b8,phi4_mini-gsm8k-s122,full_b8,4,phi4_mini,gsm8k,122,5048503,1,0,not_attempted,nan,nan,
10
+ review-control-phi4_mini-gsm8k-s122-full_b16,phi4_mini-gsm8k-s122,full_b16,1,phi4_mini,gsm8k,122,5048577,0,1,completed,35518.0,18942.0,idle
11
+ review-control-phi4_mini-gsm8k-s122-external_b8,phi4_mini-gsm8k-s122,external_b8,2,phi4_mini,gsm8k,122,5048577,0,1,completed,35518.0,18582.0,idle
12
+ review-control-phi4_mini-gsm8k-s122-external_b16,phi4_mini-gsm8k-s122,external_b16,3,phi4_mini,gsm8k,122,5048577,0,1,completed,35518.0,19234.0,idle
13
+ review-control-phi4_mini-gsm8k-s122-full_b8,phi4_mini-gsm8k-s122,full_b8,4,phi4_mini,gsm8k,122,5048577,0,1,completed,35518.0,18262.0,idle
14
+ review-control-phi4_mini-gsm8k-s123-full_b8,phi4_mini-gsm8k-s123,full_b8,1,phi4_mini,gsm8k,123,5048508,1,1,timeout_unresolved_cause,35398.0,nan,initializing
15
+ review-control-phi4_mini-gsm8k-s123-external_b16,phi4_mini-gsm8k-s123,external_b16,2,phi4_mini,gsm8k,123,5048508,1,0,not_attempted,nan,nan,
16
+ review-control-phi4_mini-gsm8k-s123-external_b8,phi4_mini-gsm8k-s123,external_b8,3,phi4_mini,gsm8k,123,5048508,1,0,not_attempted,nan,nan,
17
+ review-control-phi4_mini-gsm8k-s123-full_b16,phi4_mini-gsm8k-s123,full_b16,4,phi4_mini,gsm8k,123,5048508,1,0,not_attempted,nan,nan,
18
+ review-control-phi4_mini-gsm8k-s123-full_b8,phi4_mini-gsm8k-s123,full_b8,1,phi4_mini,gsm8k,123,5048576,0,1,completed,35518.0,16470.0,idle
19
+ review-control-phi4_mini-gsm8k-s123-external_b16,phi4_mini-gsm8k-s123,external_b16,2,phi4_mini,gsm8k,123,5048576,0,1,completed,35538.0,19298.0,idle
20
+ review-control-phi4_mini-gsm8k-s123-external_b8,phi4_mini-gsm8k-s123,external_b8,3,phi4_mini,gsm8k,123,5048576,0,1,completed,35518.0,16590.0,idle
21
+ review-control-phi4_mini-gsm8k-s123-full_b16,phi4_mini-gsm8k-s123,full_b16,4,phi4_mini,gsm8k,123,5048576,0,1,completed,35518.0,19344.0,idle
22
+ review-control-qwen25_3b-code_heavy_tail-s121-external_b16,qwen25_3b-code_heavy_tail-s121,external_b16,1,qwen25_3b,code_heavy_tail,121,5048525,1,1,completed,33386.0,25248.0,idle
23
+ review-control-qwen25_3b-code_heavy_tail-s121-full_b8,qwen25_3b-code_heavy_tail-s121,full_b8,2,qwen25_3b,code_heavy_tail,121,5048525,1,1,completed,33386.0,18668.0,idle
24
+ review-control-qwen25_3b-code_heavy_tail-s121-full_b16,qwen25_3b-code_heavy_tail-s121,full_b16,3,qwen25_3b,code_heavy_tail,121,5048525,1,1,completed,33386.0,25788.0,idle
25
+ review-control-qwen25_3b-code_heavy_tail-s121-external_b8,qwen25_3b-code_heavy_tail-s121,external_b8,4,qwen25_3b,code_heavy_tail,121,5048525,1,1,completed,33286.0,19180.0,idle
26
+ review-control-qwen25_3b-code_heavy_tail-s122-external_b8,qwen25_3b-code_heavy_tail-s122,external_b8,1,qwen25_3b,code_heavy_tail,122,5048560,1,1,completed,33546.0,19266.0,idle
27
+ review-control-qwen25_3b-code_heavy_tail-s122-external_b16,qwen25_3b-code_heavy_tail-s122,external_b16,2,qwen25_3b,code_heavy_tail,122,5048560,1,1,completed,33386.0,25188.0,idle
28
+ review-control-qwen25_3b-code_heavy_tail-s122-full_b16,qwen25_3b-code_heavy_tail-s122,full_b16,3,qwen25_3b,code_heavy_tail,122,5048560,1,1,completed,33542.0,25304.0,idle
29
+ review-control-qwen25_3b-code_heavy_tail-s122-full_b8,qwen25_3b-code_heavy_tail-s122,full_b8,4,qwen25_3b,code_heavy_tail,122,5048560,1,1,completed,33410.0,18328.0,idle
30
+ review-control-qwen25_3b-code_heavy_tail-s123-full_b16,qwen25_3b-code_heavy_tail-s123,full_b16,1,qwen25_3b,code_heavy_tail,123,5048501,1,1,completed,33386.0,26068.0,idle
31
+ review-control-qwen25_3b-code_heavy_tail-s123-full_b8,qwen25_3b-code_heavy_tail-s123,full_b8,2,qwen25_3b,code_heavy_tail,123,5048501,1,1,completed,33386.0,18988.0,idle
32
+ review-control-qwen25_3b-code_heavy_tail-s123-external_b8,qwen25_3b-code_heavy_tail-s123,external_b8,3,qwen25_3b,code_heavy_tail,123,5048501,1,1,completed,33386.0,18848.0,idle
33
+ review-control-qwen25_3b-code_heavy_tail-s123-external_b16,qwen25_3b-code_heavy_tail-s123,external_b16,4,qwen25_3b,code_heavy_tail,123,5048501,1,1,completed,33288.0,26148.0,idle
tables/review_revision/control-provenance.csv ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ pair_id,job_id,periods,gpu_count,source_head,allocation_manifest,provenance_valid
2
+ phi4_mini-gsm8k-s121,5048502,4,2,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s121/pair-5048502.json,1
3
+ phi4_mini-gsm8k-s122,5048577,4,2,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s122/pair-5048577.json,1
4
+ phi4_mini-gsm8k-s123,5048576,4,2,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s123/pair-5048576.json,1
5
+ qwen25_3b-code_heavy_tail-s121,5048525,4,2,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s121/pair-5048525.json,1
6
+ qwen25_3b-code_heavy_tail-s122,5048560,4,2,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s122/pair-5048560.json,1
7
+ qwen25_3b-code_heavy_tail-s123,5048501,4,2,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s123/pair-5048501.json,1
tables/review_revision/control-realized-work.csv ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_id,pair_id,condition,model_family,dataset,training_seed,initial_allocation,success,logged_steps,maximum_logged_step,original_train_examples,selected_before_filter,train_examples_after_filter,filtered_train_examples,positive_gradient_steps,steps_with_gradient_metric,cap_reaching_batches,batches_with_length_metrics,fixed_order_batches_per_epoch,fixed_order_distinct_train_examples_derived,fixed_order_wraps_derived,prompt_mean_over_logged_batches,prompt_maximum_observed,prompt_minimum_observed,response_mean_over_logged_batches,response_maximum_observed,response_minimum_observed
2
+ review-control-phi4_mini-gsm8k-s121-external_b8,phi4_mini-gsm8k-s121,external_b8,phi4_mini,gsm8k,121,1,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,221.65625,505.0,79.0
3
+ review-control-phi4_mini-gsm8k-s121-full_b16,phi4_mini-gsm8k-s121,full_b16,phi4_mini,gsm8k,121,1,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,231.46875,625.0,90.0
4
+ review-control-phi4_mini-gsm8k-s121-full_b8,phi4_mini-gsm8k-s121,full_b8,phi4_mini,gsm8k,121,1,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,213.40625,403.0,71.0
5
+ review-control-phi4_mini-gsm8k-s121-external_b16,phi4_mini-gsm8k-s121,external_b16,phi4_mini,gsm8k,121,1,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,227.71875,414.0,62.0
6
+ review-control-phi4_mini-gsm8k-s122-full_b16,phi4_mini-gsm8k-s122,full_b16,phi4_mini,gsm8k,122,0,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,212.40625,442.0,93.0
7
+ review-control-phi4_mini-gsm8k-s122-external_b8,phi4_mini-gsm8k-s122,external_b8,phi4_mini,gsm8k,122,0,1,1,1,7473,32,32,0,1,1,1,1,1,32,0,78.09375,121.0,43.0,266.828125,2048.0,77.0
8
+ review-control-phi4_mini-gsm8k-s122-external_b16,phi4_mini-gsm8k-s122,external_b16,phi4_mini,gsm8k,122,0,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,219.40625,455.0,86.0
9
+ review-control-phi4_mini-gsm8k-s122-full_b8,phi4_mini-gsm8k-s122,full_b8,phi4_mini,gsm8k,122,0,1,1,1,7473,32,32,0,1,1,1,1,1,32,0,78.09375,121.0,43.0,275.015625,2048.0,84.0
10
+ review-control-phi4_mini-gsm8k-s123-full_b8,phi4_mini-gsm8k-s123,full_b8,phi4_mini,gsm8k,123,0,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,217.46875,363.0,86.0
11
+ review-control-phi4_mini-gsm8k-s123-external_b16,phi4_mini-gsm8k-s123,external_b16,phi4_mini,gsm8k,123,0,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,230.265625,456.0,73.0
12
+ review-control-phi4_mini-gsm8k-s123-external_b8,phi4_mini-gsm8k-s123,external_b8,phi4_mini,gsm8k,123,0,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,209.46875,497.0,78.0
13
+ review-control-phi4_mini-gsm8k-s123-full_b16,phi4_mini-gsm8k-s123,full_b16,phi4_mini,gsm8k,123,0,1,1,1,7473,32,32,0,1,1,0,1,1,32,0,78.09375,121.0,43.0,220.015625,503.0,74.0
14
+ review-control-qwen25_3b-code_heavy_tail-s121-external_b16,qwen25_3b-code_heavy_tail-s121,external_b16,qwen25_3b,code_heavy_tail,121,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,325.703125,1024.0,105.0
15
+ review-control-qwen25_3b-code_heavy_tail-s121-full_b8,qwen25_3b-code_heavy_tail-s121,full_b8,qwen25_3b,code_heavy_tail,121,1,1,1,1,64,32,32,0,1,1,0,1,1,32,0,720.1875,772.0,674.0,346.34375,728.0,141.0
16
+ review-control-qwen25_3b-code_heavy_tail-s121-full_b16,qwen25_3b-code_heavy_tail-s121,full_b16,qwen25_3b,code_heavy_tail,121,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,350.859375,1024.0,55.0
17
+ review-control-qwen25_3b-code_heavy_tail-s121-external_b8,qwen25_3b-code_heavy_tail-s121,external_b8,qwen25_3b,code_heavy_tail,121,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,356.859375,1024.0,92.0
18
+ review-control-qwen25_3b-code_heavy_tail-s122-external_b8,qwen25_3b-code_heavy_tail-s122,external_b8,qwen25_3b,code_heavy_tail,122,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,372.859375,1024.0,71.0
19
+ review-control-qwen25_3b-code_heavy_tail-s122-external_b16,qwen25_3b-code_heavy_tail-s122,external_b16,qwen25_3b,code_heavy_tail,122,1,1,1,1,64,32,32,0,1,1,0,1,1,32,0,720.1875,772.0,674.0,333.015625,739.0,88.0
20
+ review-control-qwen25_3b-code_heavy_tail-s122-full_b16,qwen25_3b-code_heavy_tail-s122,full_b16,qwen25_3b,code_heavy_tail,122,1,1,1,1,64,32,32,0,1,1,0,1,1,32,0,720.1875,772.0,674.0,330.03125,990.0,76.0
21
+ review-control-qwen25_3b-code_heavy_tail-s122-full_b8,qwen25_3b-code_heavy_tail-s122,full_b8,qwen25_3b,code_heavy_tail,122,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,330.296875,1024.0,74.0
22
+ review-control-qwen25_3b-code_heavy_tail-s123-full_b16,qwen25_3b-code_heavy_tail-s123,full_b16,qwen25_3b,code_heavy_tail,123,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,364.3125,1024.0,75.0
23
+ review-control-qwen25_3b-code_heavy_tail-s123-full_b8,qwen25_3b-code_heavy_tail-s123,full_b8,qwen25_3b,code_heavy_tail,123,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,354.65625,1024.0,123.0
24
+ review-control-qwen25_3b-code_heavy_tail-s123-external_b8,qwen25_3b-code_heavy_tail-s123,external_b8,qwen25_3b,code_heavy_tail,123,1,1,1,1,64,32,32,0,1,1,0,1,1,32,0,720.1875,772.0,674.0,333.796875,760.0,85.0
25
+ review-control-qwen25_3b-code_heavy_tail-s123-external_b16,qwen25_3b-code_heavy_tail-s123,external_b16,qwen25_3b,code_heavy_tail,123,1,1,1,1,64,32,32,0,1,1,1,1,1,32,0,720.1875,772.0,674.0,385.328125,1024.0,123.0
tables/review_revision/control-summary.json ADDED
@@ -0,0 +1,134 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "design": "prospective_review_control_with_disclosed_bounded_extension",
3
+ "planned_slots": 24,
4
+ "observed_complete_block_slots": 24,
5
+ "completed": 24,
6
+ "within_margin": 24,
7
+ "attempts": 26,
8
+ "allocation_attempts": 8,
9
+ "cases": [
10
+ {
11
+ "model_family": "phi4_mini",
12
+ "dataset": "gsm8k",
13
+ "blocks": 3,
14
+ "actor_update_external_batch_effect_mib_observed_blocks": 3,
15
+ "actor_update_external_batch_effect_mib_mean": 1966.0,
16
+ "actor_update_external_batch_effect_mib_min": 652.0,
17
+ "actor_update_external_batch_effect_mib_max": 2708.0,
18
+ "actor_update_full_batch_effect_mib_observed_blocks": 3,
19
+ "actor_update_full_batch_effect_mib_mean": 2287.3333333333335,
20
+ "actor_update_full_batch_effect_mib_min": 680.0,
21
+ "actor_update_full_batch_effect_mib_max": 3308.0,
22
+ "actor_update_instrumentation_interaction_mib_observed_blocks": 3,
23
+ "actor_update_instrumentation_interaction_mib_mean": 321.3333333333333,
24
+ "actor_update_instrumentation_interaction_mib_min": 28.0,
25
+ "actor_update_instrumentation_interaction_mib_max": 770.0,
26
+ "rollout_external_batch_effect_mib_observed_blocks": 3,
27
+ "rollout_external_batch_effect_mib_mean": -10.666666666666666,
28
+ "rollout_external_batch_effect_mib_min": -26.0,
29
+ "rollout_external_batch_effect_mib_max": 2.0,
30
+ "rollout_full_batch_effect_mib_observed_blocks": 3,
31
+ "rollout_full_batch_effect_mib_mean": 4.0,
32
+ "rollout_full_batch_effect_mib_min": -6.0,
33
+ "rollout_full_batch_effect_mib_max": 22.0,
34
+ "rollout_instrumentation_interaction_mib_observed_blocks": 3,
35
+ "rollout_instrumentation_interaction_mib_mean": 14.666666666666666,
36
+ "rollout_instrumentation_interaction_mib_min": 2.0,
37
+ "rollout_instrumentation_interaction_mib_max": 22.0,
38
+ "weight_sync_external_batch_effect_mib_observed_blocks": 3,
39
+ "weight_sync_external_batch_effect_mib_mean": 6.666666666666667,
40
+ "weight_sync_external_batch_effect_mib_min": 0.0,
41
+ "weight_sync_external_batch_effect_mib_max": 20.0,
42
+ "weight_sync_full_batch_effect_mib_observed_blocks": 3,
43
+ "weight_sync_full_batch_effect_mib_mean": -334.6666666666667,
44
+ "weight_sync_full_batch_effect_mib_min": -1004.0,
45
+ "weight_sync_full_batch_effect_mib_max": 0.0,
46
+ "weight_sync_instrumentation_interaction_mib_observed_blocks": 3,
47
+ "weight_sync_instrumentation_interaction_mib_mean": -341.3333333333333,
48
+ "weight_sync_instrumentation_interaction_mib_min": -1004.0,
49
+ "weight_sync_instrumentation_interaction_mib_max": 0.0,
50
+ "whole_run_external_batch_effect_mib_observed_blocks": 3,
51
+ "whole_run_external_batch_effect_mib_mean": 0.0,
52
+ "whole_run_external_batch_effect_mib_min": -20.0,
53
+ "whole_run_external_batch_effect_mib_max": 20.0,
54
+ "whole_run_full_batch_effect_mib_observed_blocks": 3,
55
+ "whole_run_full_batch_effect_mib_mean": 0.0,
56
+ "whole_run_full_batch_effect_mib_min": 0.0,
57
+ "whole_run_full_batch_effect_mib_max": 0.0,
58
+ "whole_run_instrumentation_interaction_mib_observed_blocks": 3,
59
+ "whole_run_instrumentation_interaction_mib_mean": 0.0,
60
+ "whole_run_instrumentation_interaction_mib_min": -20.0,
61
+ "whole_run_instrumentation_interaction_mib_max": 20.0
62
+ },
63
+ {
64
+ "model_family": "qwen25_3b",
65
+ "dataset": "code_heavy_tail",
66
+ "blocks": 3,
67
+ "actor_update_external_batch_effect_mib_observed_blocks": 3,
68
+ "actor_update_external_batch_effect_mib_mean": 6430.0,
69
+ "actor_update_external_batch_effect_mib_min": 5922.0,
70
+ "actor_update_external_batch_effect_mib_max": 7300.0,
71
+ "actor_update_full_batch_effect_mib_observed_blocks": 3,
72
+ "actor_update_full_batch_effect_mib_mean": 7058.666666666667,
73
+ "actor_update_full_batch_effect_mib_min": 6976.0,
74
+ "actor_update_full_batch_effect_mib_max": 7120.0,
75
+ "actor_update_instrumentation_interaction_mib_observed_blocks": 3,
76
+ "actor_update_instrumentation_interaction_mib_mean": 628.6666666666666,
77
+ "actor_update_instrumentation_interaction_mib_min": -220.0,
78
+ "actor_update_instrumentation_interaction_mib_max": 1054.0,
79
+ "rollout_external_batch_effect_mib_observed_blocks": 3,
80
+ "rollout_external_batch_effect_mib_mean": -52.666666666666664,
81
+ "rollout_external_batch_effect_mib_min": -160.0,
82
+ "rollout_external_batch_effect_mib_max": 100.0,
83
+ "rollout_full_batch_effect_mib_observed_blocks": 3,
84
+ "rollout_full_batch_effect_mib_mean": 44.0,
85
+ "rollout_full_batch_effect_mib_min": 0.0,
86
+ "rollout_full_batch_effect_mib_max": 132.0,
87
+ "rollout_instrumentation_interaction_mib_observed_blocks": 3,
88
+ "rollout_instrumentation_interaction_mib_mean": 96.66666666666667,
89
+ "rollout_instrumentation_interaction_mib_min": -100.0,
90
+ "rollout_instrumentation_interaction_mib_max": 292.0,
91
+ "weight_sync_external_batch_effect_mib_observed_blocks": 3,
92
+ "weight_sync_external_batch_effect_mib_mean": -482.0,
93
+ "weight_sync_external_batch_effect_mib_min": -1446.0,
94
+ "weight_sync_external_batch_effect_mib_max": 0.0,
95
+ "weight_sync_full_batch_effect_mib_observed_blocks": 3,
96
+ "weight_sync_full_batch_effect_mib_mean": 321.3333333333333,
97
+ "weight_sync_full_batch_effect_mib_min": 0.0,
98
+ "weight_sync_full_batch_effect_mib_max": 964.0,
99
+ "weight_sync_instrumentation_interaction_mib_observed_blocks": 3,
100
+ "weight_sync_instrumentation_interaction_mib_mean": 803.3333333333334,
101
+ "weight_sync_instrumentation_interaction_mib_min": 0.0,
102
+ "weight_sync_instrumentation_interaction_mib_max": 1446.0,
103
+ "whole_run_external_batch_effect_mib_observed_blocks": 3,
104
+ "whole_run_external_batch_effect_mib_mean": -52.666666666666664,
105
+ "whole_run_external_batch_effect_mib_min": -160.0,
106
+ "whole_run_external_batch_effect_mib_max": 100.0,
107
+ "whole_run_full_batch_effect_mib_observed_blocks": 3,
108
+ "whole_run_full_batch_effect_mib_mean": 44.0,
109
+ "whole_run_full_batch_effect_mib_min": 0.0,
110
+ "whole_run_full_batch_effect_mib_max": 132.0,
111
+ "whole_run_instrumentation_interaction_mib_observed_blocks": 3,
112
+ "whole_run_instrumentation_interaction_mib_mean": 96.66666666666667,
113
+ "whole_run_instrumentation_interaction_mib_min": -100.0,
114
+ "whole_run_instrumentation_interaction_mib_max": 292.0
115
+ }
116
+ ],
117
+ "initial_attempt_outcomes": {
118
+ "accelerator_argument_error_then_noncompletion": 1,
119
+ "completed": 16,
120
+ "not_attempted": 6,
121
+ "timeout_unresolved_cause": 1
122
+ },
123
+ "additional_attempt_outcomes": {
124
+ "accelerator_argument_error_then_noncompletion": 0,
125
+ "completed": 8,
126
+ "not_attempted": 0,
127
+ "timeout_unresolved_cause": 0
128
+ },
129
+ "matrix_sha256": "21ab989275d4ecfd1279ee10545b70e9c00c172c7e13f00f426ab39d4e1ef7f6",
130
+ "source_heads": [
131
+ "018992850a9514b6dfd0d81a8060d7f8a98a89fa"
132
+ ],
133
+ "case_summary_is_descriptive_not_population_equivalence": true
134
+ }
tables/review_revision/control-trials.csv ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_id,run_group,study_stage,configuration_level,model_family,algorithm,dataset,model,data_dir,gpu_count,rollout_tp_size,actor_micro_batch,rollout_logprob_micro_batch,ref_logprob_micro_batch,vllm_gpu_memory_utilization,parameter_offload,optimizer_offload,free_cache_engine,max_prompt_length,max_response_length,max_model_len,max_num_seqs,rollout_n,train_batch_size,train_max_samples,val_max_samples,training_seed,total_training_steps,gpu_monitor_interval_ms,workload_name,custom_reward_function_path,custom_reward_function_name,revision_study,pair_id,condition,period,allocator_trace_enabled,allocator_trace_sync,resume_mode,save_freq,test_freq,val_before_train,max_actor_ckpt_to_keep,success,safe,peak_gpu_memory_mib,dominant_phase,terminal_phase,failure_kind,elapsed_seconds,job_id,artifact_path,phase_memory_csv,gpu_telemetry_csv,environment_json,run_log,allocator_trace_dir,observed_positive_steps,completed_requested_run,requested_steps,overall_peak_gpu_memory_mib,first_step_peak_mib,last_step_peak_mib,peak_drift_mib,phase_peak_initializing_mib,phase_peak_rollout_mib,phase_peak_reference_logprob_mib,phase_peak_actor_update_mib,phase_peak_weight_sync_mib,phase_peak_idle_mib,phase_steps_initializing,phase_steps_rollout,phase_steps_reference_logprob,phase_steps_actor_update,phase_steps_weight_sync,phase_steps_idle,maximum_later_step_increase_mib,per_gpu_peak_mib,risk_tier,monitor_interval_ms,source_head,allocation_manifest,allocation_hostname,initial_allocation,whole_run_peak_mib,actor_update_peak_mib,actor_update_sampled_seconds,rollout_peak_mib,rollout_sampled_seconds,weight_sync_peak_mib,weight_sync_sampled_seconds
2
+ review-control-phi4_mini-gsm8k-s121-external_b8,review_instrumentation_batch,prospective_review_control,external_b8,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,8,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,121,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s121,external_b8,1,False,1,disable,-1,-1,False,1,1,1,35538.0,idle,idle,,247,5048502,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b8/trial-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b8/attempt-5048502/phase-memory-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b8/attempt-5048502/gpu-telemetry-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b8/attempt-5048502/environment-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b8/attempt-5048502/training-5048502.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b8/attempt-5048502/allocator-traces,1,1,1,35538.0,35538.0,35538.0,0.0,35398.0,35340.0,14172.0,16824.0,35518.0,35538.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35538.0}",external_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s121/pair-5048502.json,acn22.karolina.it4i.cz,1,35538.0,16824.0,5.300004550000001,35340.0,21.700003166000016,35518.0,13.699960591999995
3
+ review-control-phi4_mini-gsm8k-s121-full_b16,review_instrumentation_batch,prospective_review_control,full_b16,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,16,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,121,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s121,full_b16,2,True,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,200,5048502,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b16/trial-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b16/attempt-5048502/phase-memory-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b16/attempt-5048502/gpu-telemetry-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b16/attempt-5048502/environment-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b16/attempt-5048502/training-5048502.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b16/attempt-5048502/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35364.0,14394.0,19760.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",full_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s121/pair-5048502.json,acn22.karolina.it4i.cz,1,35518.0,19760.0,2.799995952,35364.0,23.79999381200001,35518.0,13.800005531999997
4
+ review-control-phi4_mini-gsm8k-s121-full_b8,review_instrumentation_batch,prospective_review_control,full_b8,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,8,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,121,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s121,full_b8,3,True,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,192,5048502,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b8/trial-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b8/attempt-5048502/phase-memory-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b8/attempt-5048502/gpu-telemetry-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b8/attempt-5048502/environment-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b8/attempt-5048502/training-5048502.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-full_b8/attempt-5048502/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35342.0,14012.0,16452.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",full_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s121/pair-5048502.json,acn22.karolina.it4i.cz,1,35518.0,16452.0,3.999996655,35342.0,21.599991018000004,35518.0,11.610345091999994
5
+ review-control-phi4_mini-gsm8k-s121-external_b16,review_instrumentation_batch,prospective_review_control,external_b16,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,16,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,121,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s121,external_b16,4,False,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,189,5048502,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b16/trial-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b16/attempt-5048502/phase-memory-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b16/attempt-5048502/gpu-telemetry-5048502.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b16/attempt-5048502/environment-5048502.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b16/attempt-5048502/training-5048502.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s121-external_b16/attempt-5048502/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35342.0,14112.0,19362.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",external_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s121/pair-5048502.json,acn22.karolina.it4i.cz,1,35518.0,19362.0,2.7000078910000003,35342.0,21.999995785999992,35518.0,12.39999545
6
+ review-control-phi4_mini-gsm8k-s122-full_b16,review_instrumentation_batch,prospective_review_control,full_b16,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,16,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,122,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s122,full_b16,1,True,1,disable,-1,-1,False,1,1,1,35518.0,idle,idle,,255,5048577,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b16/trial-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b16/attempt-5048577/phase-memory-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b16/attempt-5048577/gpu-telemetry-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b16/attempt-5048577/environment-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b16/attempt-5048577/training-5048577.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b16/attempt-5048577/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35310.0,13872.0,18942.0,34514.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",full_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s122/pair-5048577.json,acn06.karolina.it4i.cz,0,35518.0,18942.0,3.3999977130000003,35310.0,31.999997785999994,34514.0,13.900002691000008
7
+ review-control-phi4_mini-gsm8k-s122-external_b8,review_instrumentation_batch,prospective_review_control,external_b8,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,8,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,122,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s122,external_b8,2,False,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,208,5048577,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b8/trial-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b8/attempt-5048577/phase-memory-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b8/attempt-5048577/gpu-telemetry-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b8/attempt-5048577/environment-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b8/attempt-5048577/training-5048577.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b8/attempt-5048577/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35340.0,15452.0,18582.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",external_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s122/pair-5048577.json,acn06.karolina.it4i.cz,0,35518.0,18582.0,4.200008011,35340.0,39.699990246,35518.0,14.104059469000005
8
+ review-control-phi4_mini-gsm8k-s122-external_b16,review_instrumentation_batch,prospective_review_control,external_b16,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,16,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,122,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s122,external_b16,3,False,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,207,5048577,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b16/trial-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b16/attempt-5048577/phase-memory-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b16/attempt-5048577/gpu-telemetry-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b16/attempt-5048577/environment-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b16/attempt-5048577/training-5048577.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-external_b16/attempt-5048577/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35332.0,14112.0,19234.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",external_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s122/pair-5048577.json,acn06.karolina.it4i.cz,0,35518.0,19234.0,4.199992105,35332.0,30.399998404000012,35518.0,14.300007439999991
9
+ review-control-phi4_mini-gsm8k-s122-full_b8,review_instrumentation_batch,prospective_review_control,full_b8,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,8,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,122,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s122,full_b8,4,True,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,220,5048577,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b8/trial-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b8/attempt-5048577/phase-memory-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b8/attempt-5048577/gpu-telemetry-5048577.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b8/attempt-5048577/environment-5048577.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b8/attempt-5048577/training-5048577.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s122-full_b8/attempt-5048577/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35316.0,15152.0,18262.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",full_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s122/pair-5048577.json,acn06.karolina.it4i.cz,0,35518.0,18262.0,4.099998761999999,35316.0,42.19999057699999,35518.0,13.900005243000004
10
+ review-control-phi4_mini-gsm8k-s123-full_b8,review_instrumentation_batch,prospective_review_control,full_b8,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,8,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,123,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s123,full_b8,1,True,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,251,5048576,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b8/trial-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b8/attempt-5048576/phase-memory-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b8/attempt-5048576/gpu-telemetry-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b8/attempt-5048576/environment-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b8/attempt-5048576/training-5048576.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b8/attempt-5048576/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35346.0,13992.0,16470.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",full_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s123/pair-5048576.json,acn06.karolina.it4i.cz,0,35518.0,16470.0,4.799997945,35346.0,25.699998282999996,35518.0,15.500006833
11
+ review-control-phi4_mini-gsm8k-s123-external_b16,review_instrumentation_batch,prospective_review_control,external_b16,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,16,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,123,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s123,external_b16,2,False,1,disable,-1,-1,False,1,1,1,35538.0,weight_sync,idle,,201,5048576,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b16/trial-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b16/attempt-5048576/phase-memory-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b16/attempt-5048576/gpu-telemetry-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b16/attempt-5048576/environment-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b16/attempt-5048576/training-5048576.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b16/attempt-5048576/allocator-traces,1,1,1,35538.0,35538.0,35538.0,0.0,35398.0,35314.0,14012.0,19298.0,35538.0,35538.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35538.0}",external_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s123/pair-5048576.json,acn06.karolina.it4i.cz,0,35538.0,19298.0,2.700003677,35314.0,28.499995345000006,35538.0,13.900006801999996
12
+ review-control-phi4_mini-gsm8k-s123-external_b8,review_instrumentation_batch,prospective_review_control,external_b8,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,8,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,123,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s123,external_b8,3,False,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,196,5048576,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b8/trial-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b8/attempt-5048576/phase-memory-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b8/attempt-5048576/gpu-telemetry-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b8/attempt-5048576/environment-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b8/attempt-5048576/training-5048576.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-external_b8/attempt-5048576/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35340.0,14092.0,16590.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",external_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s123/pair-5048576.json,acn06.karolina.it4i.cz,0,35518.0,16590.0,4.000010048000001,35340.0,25.000001097,35518.0,13.699997214000005
13
+ review-control-phi4_mini-gsm8k-s123-full_b16,review_instrumentation_batch,prospective_review_control,full_b16,phi4_mini,grpo,gsm8k,microsoft/Phi-4-mini-instruct,/mnt/proj3/open-35-44/verl/rl/data/gsm8k,2,1,16,4,4,0.6,False,False,True,1024,2048,3072,64,2,32,32,32,123,1,100,,,,review_instrumentation_batch,phi4_mini-gsm8k-s123,full_b16,4,True,1,disable,-1,-1,False,1,1,1,35518.0,weight_sync,idle,,193,5048576,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b16/trial-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b16/attempt-5048576/phase-memory-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b16/attempt-5048576/gpu-telemetry-5048576.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b16/attempt-5048576/environment-5048576.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b16/attempt-5048576/training-5048576.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-phi4_mini-gsm8k-s123-full_b16/attempt-5048576/allocator-traces,1,1,1,35518.0,35518.0,35518.0,0.0,35398.0,35342.0,14212.0,19344.0,35518.0,35518.0,0,1,1,1,1,1,nan,"{""0"": 35518.0, ""1"": 35518.0}",full_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/phi4_mini-gsm8k-s123/pair-5048576.json,acn06.karolina.it4i.cz,0,35518.0,19344.0,2.6000054610000003,35342.0,25.89999153299999,35518.0,14.099997426
14
+ review-control-qwen25_3b-code_heavy_tail-s121-external_b16,review_instrumentation_batch,prospective_review_control,external_b16,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,16,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,121,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s121,external_b16,1,False,1,disable,-1,-1,False,1,1,1,33386.0,rollout,idle,,225,5048525,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b16/trial-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b16/attempt-5048525/phase-memory-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b16/attempt-5048525/gpu-telemetry-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b16/attempt-5048525/environment-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b16/attempt-5048525/training-5048525.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b16/attempt-5048525/allocator-traces,1,1,1,33386.0,33386.0,33386.0,0.0,32674.0,33386.0,13764.0,25248.0,31662.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33386.0, ""1"": 33386.0}",external_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s121/pair-5048525.json,acn22.karolina.it4i.cz,1,33386.0,25248.0,5.900003261000001,33386.0,27.799993575999984,31662.0,9.799989798000002
15
+ review-control-qwen25_3b-code_heavy_tail-s121-full_b8,review_instrumentation_batch,prospective_review_control,full_b8,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,8,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,121,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s121,full_b8,2,True,1,disable,-1,-1,False,1,1,1,33386.0,rollout,idle,,197,5048525,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b8/trial-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b8/attempt-5048525/phase-memory-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b8/attempt-5048525/gpu-telemetry-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b8/attempt-5048525/environment-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b8/attempt-5048525/training-5048525.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b8/attempt-5048525/allocator-traces,1,1,1,33386.0,33386.0,33386.0,0.0,32674.0,33386.0,13644.0,18668.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33386.0, ""1"": 33312.0}",full_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s121/pair-5048525.json,acn22.karolina.it4i.cz,1,33386.0,18668.0,6.899995111999999,33386.0,24.299992830999997,33108.0,9.400000962000002
16
+ review-control-qwen25_3b-code_heavy_tail-s121-full_b16,review_instrumentation_batch,prospective_review_control,full_b16,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,16,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,121,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s121,full_b16,3,True,1,disable,-1,-1,False,1,1,1,33386.0,rollout,idle,,198,5048525,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b16/trial-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b16/attempt-5048525/phase-memory-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b16/attempt-5048525/gpu-telemetry-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b16/attempt-5048525/environment-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b16/attempt-5048525/training-5048525.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-full_b16/attempt-5048525/allocator-traces,1,1,1,33386.0,33386.0,33386.0,0.0,32674.0,33386.0,13804.0,25788.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33386.0, ""1"": 33294.0}",full_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s121/pair-5048525.json,acn22.karolina.it4i.cz,1,33386.0,25788.0,5.899994781999999,33386.0,27.899987331999988,33108.0,10.000044684000006
17
+ review-control-qwen25_3b-code_heavy_tail-s121-external_b8,review_instrumentation_batch,prospective_review_control,external_b8,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,8,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,121,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s121,external_b8,4,False,1,disable,-1,-1,False,1,1,1,33286.0,rollout,idle,,212,5048525,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b8/trial-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b8/attempt-5048525/phase-memory-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b8/attempt-5048525/gpu-telemetry-5048525.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b8/attempt-5048525/environment-5048525.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b8/attempt-5048525/training-5048525.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s121-external_b8/attempt-5048525/allocator-traces,1,1,1,33286.0,33286.0,33286.0,0.0,32674.0,33286.0,14024.0,19180.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33286.0, ""1"": 33286.0}",external_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s121/pair-5048525.json,acn22.karolina.it4i.cz,1,33286.0,19180.0,12.399999497000008,33286.0,28.09999576099999,33108.0,10.899999511000008
18
+ review-control-qwen25_3b-code_heavy_tail-s122-external_b8,review_instrumentation_batch,prospective_review_control,external_b8,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,8,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,122,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s122,external_b8,1,False,1,disable,-1,-1,False,1,1,1,33546.0,rollout,idle,,201,5048560,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b8/trial-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b8/attempt-5048560/phase-memory-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b8/attempt-5048560/gpu-telemetry-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b8/attempt-5048560/environment-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b8/attempt-5048560/training-5048560.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b8/attempt-5048560/allocator-traces,1,1,1,33546.0,33546.0,33546.0,0.0,32674.0,33546.0,14084.0,19266.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33324.0, ""1"": 33546.0}",external_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s122/pair-5048560.json,acn22.karolina.it4i.cz,1,33546.0,19266.0,6.999995820999998,33546.0,28.199995179000027,33108.0,10.600009411
19
+ review-control-qwen25_3b-code_heavy_tail-s122-external_b16,review_instrumentation_batch,prospective_review_control,external_b16,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,16,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,122,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s122,external_b16,2,False,1,disable,-1,-1,False,1,1,1,33386.0,rollout,idle,,201,5048560,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b16/trial-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b16/attempt-5048560/phase-memory-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b16/attempt-5048560/gpu-telemetry-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b16/attempt-5048560/environment-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b16/attempt-5048560/training-5048560.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-external_b16/attempt-5048560/allocator-traces,1,1,1,33386.0,33386.0,33386.0,0.0,32674.0,33386.0,13424.0,25188.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33386.0, ""1"": 33386.0}",external_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s122/pair-5048560.json,acn22.karolina.it4i.cz,1,33386.0,25188.0,5.900012272000001,33386.0,28.799995788,33108.0,10.299984736000008
20
+ review-control-qwen25_3b-code_heavy_tail-s122-full_b16,review_instrumentation_batch,prospective_review_control,full_b16,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,16,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,122,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s122,full_b16,3,True,1,disable,-1,-1,False,1,1,1,33542.0,rollout,idle,,202,5048560,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b16/trial-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b16/attempt-5048560/phase-memory-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b16/attempt-5048560/gpu-telemetry-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b16/attempt-5048560/environment-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b16/attempt-5048560/training-5048560.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b16/attempt-5048560/allocator-traces,1,1,1,33542.0,33542.0,33542.0,0.0,32674.0,33542.0,13724.0,25304.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33298.0, ""1"": 33542.0}",full_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s122/pair-5048560.json,acn22.karolina.it4i.cz,1,33542.0,25304.0,5.89999471,33542.0,25.999992386,33108.0,10.600011204
21
+ review-control-qwen25_3b-code_heavy_tail-s122-full_b8,review_instrumentation_batch,prospective_review_control,full_b8,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,8,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,122,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s122,full_b8,4,True,1,disable,-1,-1,False,1,1,1,33410.0,rollout,idle,,195,5048560,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b8/trial-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b8/attempt-5048560/phase-memory-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b8/attempt-5048560/gpu-telemetry-5048560.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b8/attempt-5048560/environment-5048560.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b8/attempt-5048560/training-5048560.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s122-full_b8/attempt-5048560/allocator-traces,1,1,1,33410.0,33410.0,33410.0,0.0,32676.0,33410.0,13624.0,18328.0,32144.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33356.0, ""1"": 33410.0}",full_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s122/pair-5048560.json,acn22.karolina.it4i.cz,1,33410.0,18328.0,6.900004092999998,33410.0,27.899995177999998,32144.0,9.299991953
22
+ review-control-qwen25_3b-code_heavy_tail-s123-full_b16,review_instrumentation_batch,prospective_review_control,full_b16,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,16,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,123,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s123,full_b16,1,True,1,disable,-1,-1,False,1,1,1,33386.0,rollout,idle,,201,5048501,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b16/trial-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b16/attempt-5048501/phase-memory-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b16/attempt-5048501/gpu-telemetry-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b16/attempt-5048501/environment-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b16/attempt-5048501/training-5048501.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b16/attempt-5048501/allocator-traces,1,1,1,33386.0,33386.0,33386.0,0.0,32676.0,33386.0,14084.0,26068.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33386.0, ""1"": 33386.0}",full_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s123/pair-5048501.json,acn22.karolina.it4i.cz,1,33386.0,26068.0,7.400003674000001,33386.0,30.199998720999996,33108.0,9.400000109999999
23
+ review-control-qwen25_3b-code_heavy_tail-s123-full_b8,review_instrumentation_batch,prospective_review_control,full_b8,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,8,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,123,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s123,full_b8,2,True,1,disable,-1,-1,False,1,1,1,33386.0,rollout,idle,,209,5048501,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b8/trial-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b8/attempt-5048501/phase-memory-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b8/attempt-5048501/gpu-telemetry-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b8/attempt-5048501/environment-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b8/attempt-5048501/training-5048501.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-full_b8/attempt-5048501/allocator-traces,1,1,1,33386.0,33386.0,33386.0,0.0,32676.0,33386.0,13944.0,18988.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33386.0, ""1"": 33386.0}",full_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s123/pair-5048501.json,acn22.karolina.it4i.cz,1,33386.0,18988.0,7.100003771999999,33386.0,30.699991110000017,33108.0,10.699995154999998
24
+ review-control-qwen25_3b-code_heavy_tail-s123-external_b8,review_instrumentation_batch,prospective_review_control,external_b8,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,8,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,123,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s123,external_b8,3,False,1,disable,-1,-1,False,1,1,1,33386.0,rollout,idle,,192,5048501,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b8/trial-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b8/attempt-5048501/phase-memory-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b8/attempt-5048501/gpu-telemetry-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b8/attempt-5048501/environment-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b8/attempt-5048501/training-5048501.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b8/attempt-5048501/allocator-traces,1,1,1,33386.0,33386.0,33386.0,0.0,32674.0,33386.0,13724.0,18848.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33318.0, ""1"": 33386.0}",external_b8,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s123/pair-5048501.json,acn22.karolina.it4i.cz,1,33386.0,18848.0,7.099998215999998,33386.0,24.099991005,33108.0,10.600002063
25
+ review-control-qwen25_3b-code_heavy_tail-s123-external_b16,review_instrumentation_batch,prospective_review_control,external_b16,qwen25_3b,grpo,code_heavy_tail,Qwen/Qwen2.5-3B-Instruct,/mnt/proj3/open-35-44/verl/rl/data/codecontests/heavy_tail,2,1,16,4,4,0.6,False,False,True,1536,1024,2560,64,2,32,32,32,123,1,100,code_heavy_tail,/mnt/proj3/open-35-44/verl/rl/memory_tuner/code_shape_reward.py,compute_score,review_instrumentation_batch,qwen25_3b-code_heavy_tail-s123,external_b16,4,False,1,disable,-1,-1,False,1,1,1,33288.0,rollout,idle,,193,5048501,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b16/trial-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b16/attempt-5048501/phase-memory-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b16/attempt-5048501/gpu-telemetry-5048501.csv,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b16/attempt-5048501/environment-5048501.json,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b16/attempt-5048501/training-5048501.log,/mnt/proj3/open-35-44/verl/rl/output/review_instrumentation_batch/review-control-qwen25_3b-code_heavy_tail-s123-external_b16/attempt-5048501/allocator-traces,1,1,1,33288.0,33288.0,33288.0,0.0,32674.0,33288.0,14084.0,26148.0,33108.0,33108.0,0,1,1,1,1,1,nan,"{""0"": 33288.0, ""1"": 33262.0}",external_b16,100,018992850a9514b6dfd0d81a8060d7f8a98a89fa,output/review_instrumentation_batch/pairs/qwen25_3b-code_heavy_tail-s123/pair-5048501.json,acn22.karolina.it4i.cz,1,33288.0,26148.0,6.200007945000001,33288.0,27.99999680999998,33108.0,9.500000165000001
tables/review_revision/corpus-weighting-sensitivity.csv ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ estimand,four_workload_mean,three_corpus_equal_weight_mean,minimum_workload_effect,maximum_workload_effect
2
+ actor_main_effect_whole_run_mib,-30.61111111111111,-22.59259259259259,-95.44444444444444,0.0
3
+ reservation_main_effect_whole_run_mib,3998.0,4006.4074074074074,3881.4444444444443,4064.1111111111113
4
+ interaction_whole_run_mib,20.777777777777775,2.2222222222222197,-74.22222222222223,192.66666666666666
5
+ actor_main_effect_actor_nvml_mib,3980.0555555555557,3573.5,2072.8888888888887,6267.555555555556
6
+ reservation_main_effect_actor_nvml_mib,8.83333333333334,41.16666666666667,-343.55555555555554,173.22222222222223
7
+ interaction_actor_nvml_mib,114.66666666666666,147.0,-34.22222222222222,305.55555555555554
8
+ actor_main_effect_actor_allocated_mib,3847.7573310004336,3450.473614728009,2000.4676649305557,6072.907172309027
9
+ reservation_main_effect_actor_allocated_mib,0.6370510525173589,30.767424406828702,-309.62456597222223,133.652099609375
10
+ interaction_actor_allocated_mib,111.80247667100694,148.81673177083331,-40.21929253472222,259.77088758680554
tables/review_revision/decision-records.csv ADDED
@@ -0,0 +1,397 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ rule,donor,target,dataset,configuration_level,prediction,observed,model_family
2
+ source_model,granite33_2b,phi4_mini,code_heavy_tail,c0,within_margin,within_margin,
3
+ source_model,granite33_2b,phi4_mini,code_heavy_tail,c1,within_margin,within_margin,
4
+ source_model,granite33_2b,phi4_mini,code_heavy_tail,c2,within_margin,within_margin,
5
+ source_model,granite33_2b,phi4_mini,code_heavy_tail,c3,within_margin,above_margin,
6
+ source_model,granite33_2b,phi4_mini,code_heavy_tail,c4,above_margin,memory_failure,
7
+ source_model,granite33_2b,phi4_mini,code_heavy_tail,c5,memory_failure,memory_failure,
8
+ source_model,granite33_2b,phi4_mini,code_standard,c0,within_margin,within_margin,
9
+ source_model,granite33_2b,phi4_mini,code_standard,c1,within_margin,within_margin,
10
+ source_model,granite33_2b,phi4_mini,code_standard,c2,within_margin,within_margin,
11
+ source_model,granite33_2b,phi4_mini,code_standard,c3,within_margin,above_margin,
12
+ source_model,granite33_2b,phi4_mini,code_standard,c4,above_margin,memory_failure,
13
+ source_model,granite33_2b,phi4_mini,code_standard,c5,memory_failure,memory_failure,
14
+ source_model,granite33_2b,phi4_mini,gsm8k,c0,within_margin,within_margin,
15
+ source_model,granite33_2b,phi4_mini,gsm8k,c1,within_margin,within_margin,
16
+ source_model,granite33_2b,phi4_mini,gsm8k,c2,within_margin,within_margin,
17
+ source_model,granite33_2b,phi4_mini,gsm8k,c3,within_margin,above_margin,
18
+ source_model,granite33_2b,phi4_mini,gsm8k,c4,above_margin,memory_failure,
19
+ source_model,granite33_2b,phi4_mini,gsm8k,c5,memory_failure,memory_failure,
20
+ source_model,granite33_2b,phi4_mini,math,c0,within_margin,within_margin,
21
+ source_model,granite33_2b,phi4_mini,math,c1,within_margin,within_margin,
22
+ source_model,granite33_2b,phi4_mini,math,c2,within_margin,within_margin,
23
+ source_model,granite33_2b,phi4_mini,math,c3,within_margin,above_margin,
24
+ source_model,granite33_2b,phi4_mini,math,c4,above_margin,memory_failure,
25
+ source_model,granite33_2b,phi4_mini,math,c5,memory_failure,memory_failure,
26
+ source_model,granite33_2b,qwen25_3b,code_heavy_tail,c0,within_margin,within_margin,
27
+ source_model,granite33_2b,qwen25_3b,code_heavy_tail,c1,within_margin,within_margin,
28
+ source_model,granite33_2b,qwen25_3b,code_heavy_tail,c2,within_margin,within_margin,
29
+ source_model,granite33_2b,qwen25_3b,code_heavy_tail,c3,within_margin,within_margin,
30
+ source_model,granite33_2b,qwen25_3b,code_heavy_tail,c4,above_margin,memory_failure,
31
+ source_model,granite33_2b,qwen25_3b,code_heavy_tail,c5,memory_failure,memory_failure,
32
+ source_model,granite33_2b,qwen25_3b,code_standard,c0,within_margin,within_margin,
33
+ source_model,granite33_2b,qwen25_3b,code_standard,c1,within_margin,within_margin,
34
+ source_model,granite33_2b,qwen25_3b,code_standard,c2,within_margin,within_margin,
35
+ source_model,granite33_2b,qwen25_3b,code_standard,c3,within_margin,within_margin,
36
+ source_model,granite33_2b,qwen25_3b,code_standard,c4,above_margin,memory_failure,
37
+ source_model,granite33_2b,qwen25_3b,code_standard,c5,memory_failure,memory_failure,
38
+ source_model,granite33_2b,qwen25_3b,gsm8k,c0,within_margin,within_margin,
39
+ source_model,granite33_2b,qwen25_3b,gsm8k,c1,within_margin,within_margin,
40
+ source_model,granite33_2b,qwen25_3b,gsm8k,c2,within_margin,within_margin,
41
+ source_model,granite33_2b,qwen25_3b,gsm8k,c3,within_margin,within_margin,
42
+ source_model,granite33_2b,qwen25_3b,gsm8k,c4,above_margin,memory_failure,
43
+ source_model,granite33_2b,qwen25_3b,gsm8k,c5,memory_failure,memory_failure,
44
+ source_model,granite33_2b,qwen25_3b,math,c0,within_margin,within_margin,
45
+ source_model,granite33_2b,qwen25_3b,math,c1,within_margin,within_margin,
46
+ source_model,granite33_2b,qwen25_3b,math,c2,within_margin,within_margin,
47
+ source_model,granite33_2b,qwen25_3b,math,c3,within_margin,within_margin,
48
+ source_model,granite33_2b,qwen25_3b,math,c4,above_margin,memory_failure,
49
+ source_model,granite33_2b,qwen25_3b,math,c5,memory_failure,memory_failure,
50
+ source_model,phi4_mini,granite33_2b,code_heavy_tail,c0,within_margin,within_margin,
51
+ source_model,phi4_mini,granite33_2b,code_heavy_tail,c1,within_margin,within_margin,
52
+ source_model,phi4_mini,granite33_2b,code_heavy_tail,c2,within_margin,within_margin,
53
+ source_model,phi4_mini,granite33_2b,code_heavy_tail,c3,above_margin,within_margin,
54
+ source_model,phi4_mini,granite33_2b,code_heavy_tail,c4,memory_failure,above_margin,
55
+ source_model,phi4_mini,granite33_2b,code_heavy_tail,c5,memory_failure,memory_failure,
56
+ source_model,phi4_mini,granite33_2b,code_standard,c0,within_margin,within_margin,
57
+ source_model,phi4_mini,granite33_2b,code_standard,c1,within_margin,within_margin,
58
+ source_model,phi4_mini,granite33_2b,code_standard,c2,within_margin,within_margin,
59
+ source_model,phi4_mini,granite33_2b,code_standard,c3,above_margin,within_margin,
60
+ source_model,phi4_mini,granite33_2b,code_standard,c4,memory_failure,above_margin,
61
+ source_model,phi4_mini,granite33_2b,code_standard,c5,memory_failure,memory_failure,
62
+ source_model,phi4_mini,granite33_2b,gsm8k,c0,within_margin,within_margin,
63
+ source_model,phi4_mini,granite33_2b,gsm8k,c1,within_margin,within_margin,
64
+ source_model,phi4_mini,granite33_2b,gsm8k,c2,within_margin,within_margin,
65
+ source_model,phi4_mini,granite33_2b,gsm8k,c3,above_margin,within_margin,
66
+ source_model,phi4_mini,granite33_2b,gsm8k,c4,memory_failure,above_margin,
67
+ source_model,phi4_mini,granite33_2b,gsm8k,c5,memory_failure,memory_failure,
68
+ source_model,phi4_mini,granite33_2b,math,c0,within_margin,within_margin,
69
+ source_model,phi4_mini,granite33_2b,math,c1,within_margin,within_margin,
70
+ source_model,phi4_mini,granite33_2b,math,c2,within_margin,within_margin,
71
+ source_model,phi4_mini,granite33_2b,math,c3,above_margin,within_margin,
72
+ source_model,phi4_mini,granite33_2b,math,c4,memory_failure,above_margin,
73
+ source_model,phi4_mini,granite33_2b,math,c5,memory_failure,memory_failure,
74
+ source_model,phi4_mini,qwen25_3b,code_heavy_tail,c0,within_margin,within_margin,
75
+ source_model,phi4_mini,qwen25_3b,code_heavy_tail,c1,within_margin,within_margin,
76
+ source_model,phi4_mini,qwen25_3b,code_heavy_tail,c2,within_margin,within_margin,
77
+ source_model,phi4_mini,qwen25_3b,code_heavy_tail,c3,above_margin,within_margin,
78
+ source_model,phi4_mini,qwen25_3b,code_heavy_tail,c4,memory_failure,memory_failure,
79
+ source_model,phi4_mini,qwen25_3b,code_heavy_tail,c5,memory_failure,memory_failure,
80
+ source_model,phi4_mini,qwen25_3b,code_standard,c0,within_margin,within_margin,
81
+ source_model,phi4_mini,qwen25_3b,code_standard,c1,within_margin,within_margin,
82
+ source_model,phi4_mini,qwen25_3b,code_standard,c2,within_margin,within_margin,
83
+ source_model,phi4_mini,qwen25_3b,code_standard,c3,above_margin,within_margin,
84
+ source_model,phi4_mini,qwen25_3b,code_standard,c4,memory_failure,memory_failure,
85
+ source_model,phi4_mini,qwen25_3b,code_standard,c5,memory_failure,memory_failure,
86
+ source_model,phi4_mini,qwen25_3b,gsm8k,c0,within_margin,within_margin,
87
+ source_model,phi4_mini,qwen25_3b,gsm8k,c1,within_margin,within_margin,
88
+ source_model,phi4_mini,qwen25_3b,gsm8k,c2,within_margin,within_margin,
89
+ source_model,phi4_mini,qwen25_3b,gsm8k,c3,above_margin,within_margin,
90
+ source_model,phi4_mini,qwen25_3b,gsm8k,c4,memory_failure,memory_failure,
91
+ source_model,phi4_mini,qwen25_3b,gsm8k,c5,memory_failure,memory_failure,
92
+ source_model,phi4_mini,qwen25_3b,math,c0,within_margin,within_margin,
93
+ source_model,phi4_mini,qwen25_3b,math,c1,within_margin,within_margin,
94
+ source_model,phi4_mini,qwen25_3b,math,c2,within_margin,within_margin,
95
+ source_model,phi4_mini,qwen25_3b,math,c3,above_margin,within_margin,
96
+ source_model,phi4_mini,qwen25_3b,math,c4,memory_failure,memory_failure,
97
+ source_model,phi4_mini,qwen25_3b,math,c5,memory_failure,memory_failure,
98
+ source_model,qwen25_3b,granite33_2b,code_heavy_tail,c0,within_margin,within_margin,
99
+ source_model,qwen25_3b,granite33_2b,code_heavy_tail,c1,within_margin,within_margin,
100
+ source_model,qwen25_3b,granite33_2b,code_heavy_tail,c2,within_margin,within_margin,
101
+ source_model,qwen25_3b,granite33_2b,code_heavy_tail,c3,within_margin,within_margin,
102
+ source_model,qwen25_3b,granite33_2b,code_heavy_tail,c4,memory_failure,above_margin,
103
+ source_model,qwen25_3b,granite33_2b,code_heavy_tail,c5,memory_failure,memory_failure,
104
+ source_model,qwen25_3b,granite33_2b,code_standard,c0,within_margin,within_margin,
105
+ source_model,qwen25_3b,granite33_2b,code_standard,c1,within_margin,within_margin,
106
+ source_model,qwen25_3b,granite33_2b,code_standard,c2,within_margin,within_margin,
107
+ source_model,qwen25_3b,granite33_2b,code_standard,c3,within_margin,within_margin,
108
+ source_model,qwen25_3b,granite33_2b,code_standard,c4,memory_failure,above_margin,
109
+ source_model,qwen25_3b,granite33_2b,code_standard,c5,memory_failure,memory_failure,
110
+ source_model,qwen25_3b,granite33_2b,gsm8k,c0,within_margin,within_margin,
111
+ source_model,qwen25_3b,granite33_2b,gsm8k,c1,within_margin,within_margin,
112
+ source_model,qwen25_3b,granite33_2b,gsm8k,c2,within_margin,within_margin,
113
+ source_model,qwen25_3b,granite33_2b,gsm8k,c3,within_margin,within_margin,
114
+ source_model,qwen25_3b,granite33_2b,gsm8k,c4,memory_failure,above_margin,
115
+ source_model,qwen25_3b,granite33_2b,gsm8k,c5,memory_failure,memory_failure,
116
+ source_model,qwen25_3b,granite33_2b,math,c0,within_margin,within_margin,
117
+ source_model,qwen25_3b,granite33_2b,math,c1,within_margin,within_margin,
118
+ source_model,qwen25_3b,granite33_2b,math,c2,within_margin,within_margin,
119
+ source_model,qwen25_3b,granite33_2b,math,c3,within_margin,within_margin,
120
+ source_model,qwen25_3b,granite33_2b,math,c4,memory_failure,above_margin,
121
+ source_model,qwen25_3b,granite33_2b,math,c5,memory_failure,memory_failure,
122
+ source_model,qwen25_3b,phi4_mini,code_heavy_tail,c0,within_margin,within_margin,
123
+ source_model,qwen25_3b,phi4_mini,code_heavy_tail,c1,within_margin,within_margin,
124
+ source_model,qwen25_3b,phi4_mini,code_heavy_tail,c2,within_margin,within_margin,
125
+ source_model,qwen25_3b,phi4_mini,code_heavy_tail,c3,within_margin,above_margin,
126
+ source_model,qwen25_3b,phi4_mini,code_heavy_tail,c4,memory_failure,memory_failure,
127
+ source_model,qwen25_3b,phi4_mini,code_heavy_tail,c5,memory_failure,memory_failure,
128
+ source_model,qwen25_3b,phi4_mini,code_standard,c0,within_margin,within_margin,
129
+ source_model,qwen25_3b,phi4_mini,code_standard,c1,within_margin,within_margin,
130
+ source_model,qwen25_3b,phi4_mini,code_standard,c2,within_margin,within_margin,
131
+ source_model,qwen25_3b,phi4_mini,code_standard,c3,within_margin,above_margin,
132
+ source_model,qwen25_3b,phi4_mini,code_standard,c4,memory_failure,memory_failure,
133
+ source_model,qwen25_3b,phi4_mini,code_standard,c5,memory_failure,memory_failure,
134
+ source_model,qwen25_3b,phi4_mini,gsm8k,c0,within_margin,within_margin,
135
+ source_model,qwen25_3b,phi4_mini,gsm8k,c1,within_margin,within_margin,
136
+ source_model,qwen25_3b,phi4_mini,gsm8k,c2,within_margin,within_margin,
137
+ source_model,qwen25_3b,phi4_mini,gsm8k,c3,within_margin,above_margin,
138
+ source_model,qwen25_3b,phi4_mini,gsm8k,c4,memory_failure,memory_failure,
139
+ source_model,qwen25_3b,phi4_mini,gsm8k,c5,memory_failure,memory_failure,
140
+ source_model,qwen25_3b,phi4_mini,math,c0,within_margin,within_margin,
141
+ source_model,qwen25_3b,phi4_mini,math,c1,within_margin,within_margin,
142
+ source_model,qwen25_3b,phi4_mini,math,c2,within_margin,within_margin,
143
+ source_model,qwen25_3b,phi4_mini,math,c3,within_margin,above_margin,
144
+ source_model,qwen25_3b,phi4_mini,math,c4,memory_failure,memory_failure,
145
+ source_model,qwen25_3b,phi4_mini,math,c5,memory_failure,memory_failure,
146
+ donor_workload_lookup,code_heavy_tail,code_standard,,c0,within_margin,within_margin,granite33_2b
147
+ donor_workload_lookup,code_heavy_tail,code_standard,,c1,within_margin,within_margin,granite33_2b
148
+ donor_workload_lookup,code_heavy_tail,code_standard,,c2,within_margin,within_margin,granite33_2b
149
+ donor_workload_lookup,code_heavy_tail,code_standard,,c3,within_margin,within_margin,granite33_2b
150
+ donor_workload_lookup,code_heavy_tail,code_standard,,c4,above_margin,above_margin,granite33_2b
151
+ donor_workload_lookup,code_heavy_tail,code_standard,,c5,memory_failure,memory_failure,granite33_2b
152
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c0,within_margin,within_margin,granite33_2b
153
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c1,within_margin,within_margin,granite33_2b
154
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c2,within_margin,within_margin,granite33_2b
155
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c3,within_margin,within_margin,granite33_2b
156
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c4,above_margin,above_margin,granite33_2b
157
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c5,memory_failure,memory_failure,granite33_2b
158
+ donor_workload_lookup,code_heavy_tail,math,,c0,within_margin,within_margin,granite33_2b
159
+ donor_workload_lookup,code_heavy_tail,math,,c1,within_margin,within_margin,granite33_2b
160
+ donor_workload_lookup,code_heavy_tail,math,,c2,within_margin,within_margin,granite33_2b
161
+ donor_workload_lookup,code_heavy_tail,math,,c3,within_margin,within_margin,granite33_2b
162
+ donor_workload_lookup,code_heavy_tail,math,,c4,above_margin,above_margin,granite33_2b
163
+ donor_workload_lookup,code_heavy_tail,math,,c5,memory_failure,memory_failure,granite33_2b
164
+ donor_workload_lookup,code_standard,code_heavy_tail,,c0,within_margin,within_margin,granite33_2b
165
+ donor_workload_lookup,code_standard,code_heavy_tail,,c1,within_margin,within_margin,granite33_2b
166
+ donor_workload_lookup,code_standard,code_heavy_tail,,c2,within_margin,within_margin,granite33_2b
167
+ donor_workload_lookup,code_standard,code_heavy_tail,,c3,within_margin,within_margin,granite33_2b
168
+ donor_workload_lookup,code_standard,code_heavy_tail,,c4,above_margin,above_margin,granite33_2b
169
+ donor_workload_lookup,code_standard,code_heavy_tail,,c5,memory_failure,memory_failure,granite33_2b
170
+ donor_workload_lookup,code_standard,gsm8k,,c0,within_margin,within_margin,granite33_2b
171
+ donor_workload_lookup,code_standard,gsm8k,,c1,within_margin,within_margin,granite33_2b
172
+ donor_workload_lookup,code_standard,gsm8k,,c2,within_margin,within_margin,granite33_2b
173
+ donor_workload_lookup,code_standard,gsm8k,,c3,within_margin,within_margin,granite33_2b
174
+ donor_workload_lookup,code_standard,gsm8k,,c4,above_margin,above_margin,granite33_2b
175
+ donor_workload_lookup,code_standard,gsm8k,,c5,memory_failure,memory_failure,granite33_2b
176
+ donor_workload_lookup,code_standard,math,,c0,within_margin,within_margin,granite33_2b
177
+ donor_workload_lookup,code_standard,math,,c1,within_margin,within_margin,granite33_2b
178
+ donor_workload_lookup,code_standard,math,,c2,within_margin,within_margin,granite33_2b
179
+ donor_workload_lookup,code_standard,math,,c3,within_margin,within_margin,granite33_2b
180
+ donor_workload_lookup,code_standard,math,,c4,above_margin,above_margin,granite33_2b
181
+ donor_workload_lookup,code_standard,math,,c5,memory_failure,memory_failure,granite33_2b
182
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c0,within_margin,within_margin,granite33_2b
183
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c1,within_margin,within_margin,granite33_2b
184
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c2,within_margin,within_margin,granite33_2b
185
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c3,within_margin,within_margin,granite33_2b
186
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c4,above_margin,above_margin,granite33_2b
187
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c5,memory_failure,memory_failure,granite33_2b
188
+ donor_workload_lookup,gsm8k,code_standard,,c0,within_margin,within_margin,granite33_2b
189
+ donor_workload_lookup,gsm8k,code_standard,,c1,within_margin,within_margin,granite33_2b
190
+ donor_workload_lookup,gsm8k,code_standard,,c2,within_margin,within_margin,granite33_2b
191
+ donor_workload_lookup,gsm8k,code_standard,,c3,within_margin,within_margin,granite33_2b
192
+ donor_workload_lookup,gsm8k,code_standard,,c4,above_margin,above_margin,granite33_2b
193
+ donor_workload_lookup,gsm8k,code_standard,,c5,memory_failure,memory_failure,granite33_2b
194
+ donor_workload_lookup,gsm8k,math,,c0,within_margin,within_margin,granite33_2b
195
+ donor_workload_lookup,gsm8k,math,,c1,within_margin,within_margin,granite33_2b
196
+ donor_workload_lookup,gsm8k,math,,c2,within_margin,within_margin,granite33_2b
197
+ donor_workload_lookup,gsm8k,math,,c3,within_margin,within_margin,granite33_2b
198
+ donor_workload_lookup,gsm8k,math,,c4,above_margin,above_margin,granite33_2b
199
+ donor_workload_lookup,gsm8k,math,,c5,memory_failure,memory_failure,granite33_2b
200
+ donor_workload_lookup,math,code_heavy_tail,,c0,within_margin,within_margin,granite33_2b
201
+ donor_workload_lookup,math,code_heavy_tail,,c1,within_margin,within_margin,granite33_2b
202
+ donor_workload_lookup,math,code_heavy_tail,,c2,within_margin,within_margin,granite33_2b
203
+ donor_workload_lookup,math,code_heavy_tail,,c3,within_margin,within_margin,granite33_2b
204
+ donor_workload_lookup,math,code_heavy_tail,,c4,above_margin,above_margin,granite33_2b
205
+ donor_workload_lookup,math,code_heavy_tail,,c5,memory_failure,memory_failure,granite33_2b
206
+ donor_workload_lookup,math,code_standard,,c0,within_margin,within_margin,granite33_2b
207
+ donor_workload_lookup,math,code_standard,,c1,within_margin,within_margin,granite33_2b
208
+ donor_workload_lookup,math,code_standard,,c2,within_margin,within_margin,granite33_2b
209
+ donor_workload_lookup,math,code_standard,,c3,within_margin,within_margin,granite33_2b
210
+ donor_workload_lookup,math,code_standard,,c4,above_margin,above_margin,granite33_2b
211
+ donor_workload_lookup,math,code_standard,,c5,memory_failure,memory_failure,granite33_2b
212
+ donor_workload_lookup,math,gsm8k,,c0,within_margin,within_margin,granite33_2b
213
+ donor_workload_lookup,math,gsm8k,,c1,within_margin,within_margin,granite33_2b
214
+ donor_workload_lookup,math,gsm8k,,c2,within_margin,within_margin,granite33_2b
215
+ donor_workload_lookup,math,gsm8k,,c3,within_margin,within_margin,granite33_2b
216
+ donor_workload_lookup,math,gsm8k,,c4,above_margin,above_margin,granite33_2b
217
+ donor_workload_lookup,math,gsm8k,,c5,memory_failure,memory_failure,granite33_2b
218
+ donor_workload_lookup,code_heavy_tail,code_standard,,c0,within_margin,within_margin,phi4_mini
219
+ donor_workload_lookup,code_heavy_tail,code_standard,,c1,within_margin,within_margin,phi4_mini
220
+ donor_workload_lookup,code_heavy_tail,code_standard,,c2,within_margin,within_margin,phi4_mini
221
+ donor_workload_lookup,code_heavy_tail,code_standard,,c3,above_margin,above_margin,phi4_mini
222
+ donor_workload_lookup,code_heavy_tail,code_standard,,c4,memory_failure,memory_failure,phi4_mini
223
+ donor_workload_lookup,code_heavy_tail,code_standard,,c5,memory_failure,memory_failure,phi4_mini
224
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c0,within_margin,within_margin,phi4_mini
225
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c1,within_margin,within_margin,phi4_mini
226
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c2,within_margin,within_margin,phi4_mini
227
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c3,above_margin,above_margin,phi4_mini
228
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c4,memory_failure,memory_failure,phi4_mini
229
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c5,memory_failure,memory_failure,phi4_mini
230
+ donor_workload_lookup,code_heavy_tail,math,,c0,within_margin,within_margin,phi4_mini
231
+ donor_workload_lookup,code_heavy_tail,math,,c1,within_margin,within_margin,phi4_mini
232
+ donor_workload_lookup,code_heavy_tail,math,,c2,within_margin,within_margin,phi4_mini
233
+ donor_workload_lookup,code_heavy_tail,math,,c3,above_margin,above_margin,phi4_mini
234
+ donor_workload_lookup,code_heavy_tail,math,,c4,memory_failure,memory_failure,phi4_mini
235
+ donor_workload_lookup,code_heavy_tail,math,,c5,memory_failure,memory_failure,phi4_mini
236
+ donor_workload_lookup,code_standard,code_heavy_tail,,c0,within_margin,within_margin,phi4_mini
237
+ donor_workload_lookup,code_standard,code_heavy_tail,,c1,within_margin,within_margin,phi4_mini
238
+ donor_workload_lookup,code_standard,code_heavy_tail,,c2,within_margin,within_margin,phi4_mini
239
+ donor_workload_lookup,code_standard,code_heavy_tail,,c3,above_margin,above_margin,phi4_mini
240
+ donor_workload_lookup,code_standard,code_heavy_tail,,c4,memory_failure,memory_failure,phi4_mini
241
+ donor_workload_lookup,code_standard,code_heavy_tail,,c5,memory_failure,memory_failure,phi4_mini
242
+ donor_workload_lookup,code_standard,gsm8k,,c0,within_margin,within_margin,phi4_mini
243
+ donor_workload_lookup,code_standard,gsm8k,,c1,within_margin,within_margin,phi4_mini
244
+ donor_workload_lookup,code_standard,gsm8k,,c2,within_margin,within_margin,phi4_mini
245
+ donor_workload_lookup,code_standard,gsm8k,,c3,above_margin,above_margin,phi4_mini
246
+ donor_workload_lookup,code_standard,gsm8k,,c4,memory_failure,memory_failure,phi4_mini
247
+ donor_workload_lookup,code_standard,gsm8k,,c5,memory_failure,memory_failure,phi4_mini
248
+ donor_workload_lookup,code_standard,math,,c0,within_margin,within_margin,phi4_mini
249
+ donor_workload_lookup,code_standard,math,,c1,within_margin,within_margin,phi4_mini
250
+ donor_workload_lookup,code_standard,math,,c2,within_margin,within_margin,phi4_mini
251
+ donor_workload_lookup,code_standard,math,,c3,above_margin,above_margin,phi4_mini
252
+ donor_workload_lookup,code_standard,math,,c4,memory_failure,memory_failure,phi4_mini
253
+ donor_workload_lookup,code_standard,math,,c5,memory_failure,memory_failure,phi4_mini
254
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c0,within_margin,within_margin,phi4_mini
255
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c1,within_margin,within_margin,phi4_mini
256
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c2,within_margin,within_margin,phi4_mini
257
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c3,above_margin,above_margin,phi4_mini
258
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c4,memory_failure,memory_failure,phi4_mini
259
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c5,memory_failure,memory_failure,phi4_mini
260
+ donor_workload_lookup,gsm8k,code_standard,,c0,within_margin,within_margin,phi4_mini
261
+ donor_workload_lookup,gsm8k,code_standard,,c1,within_margin,within_margin,phi4_mini
262
+ donor_workload_lookup,gsm8k,code_standard,,c2,within_margin,within_margin,phi4_mini
263
+ donor_workload_lookup,gsm8k,code_standard,,c3,above_margin,above_margin,phi4_mini
264
+ donor_workload_lookup,gsm8k,code_standard,,c4,memory_failure,memory_failure,phi4_mini
265
+ donor_workload_lookup,gsm8k,code_standard,,c5,memory_failure,memory_failure,phi4_mini
266
+ donor_workload_lookup,gsm8k,math,,c0,within_margin,within_margin,phi4_mini
267
+ donor_workload_lookup,gsm8k,math,,c1,within_margin,within_margin,phi4_mini
268
+ donor_workload_lookup,gsm8k,math,,c2,within_margin,within_margin,phi4_mini
269
+ donor_workload_lookup,gsm8k,math,,c3,above_margin,above_margin,phi4_mini
270
+ donor_workload_lookup,gsm8k,math,,c4,memory_failure,memory_failure,phi4_mini
271
+ donor_workload_lookup,gsm8k,math,,c5,memory_failure,memory_failure,phi4_mini
272
+ donor_workload_lookup,math,code_heavy_tail,,c0,within_margin,within_margin,phi4_mini
273
+ donor_workload_lookup,math,code_heavy_tail,,c1,within_margin,within_margin,phi4_mini
274
+ donor_workload_lookup,math,code_heavy_tail,,c2,within_margin,within_margin,phi4_mini
275
+ donor_workload_lookup,math,code_heavy_tail,,c3,above_margin,above_margin,phi4_mini
276
+ donor_workload_lookup,math,code_heavy_tail,,c4,memory_failure,memory_failure,phi4_mini
277
+ donor_workload_lookup,math,code_heavy_tail,,c5,memory_failure,memory_failure,phi4_mini
278
+ donor_workload_lookup,math,code_standard,,c0,within_margin,within_margin,phi4_mini
279
+ donor_workload_lookup,math,code_standard,,c1,within_margin,within_margin,phi4_mini
280
+ donor_workload_lookup,math,code_standard,,c2,within_margin,within_margin,phi4_mini
281
+ donor_workload_lookup,math,code_standard,,c3,above_margin,above_margin,phi4_mini
282
+ donor_workload_lookup,math,code_standard,,c4,memory_failure,memory_failure,phi4_mini
283
+ donor_workload_lookup,math,code_standard,,c5,memory_failure,memory_failure,phi4_mini
284
+ donor_workload_lookup,math,gsm8k,,c0,within_margin,within_margin,phi4_mini
285
+ donor_workload_lookup,math,gsm8k,,c1,within_margin,within_margin,phi4_mini
286
+ donor_workload_lookup,math,gsm8k,,c2,within_margin,within_margin,phi4_mini
287
+ donor_workload_lookup,math,gsm8k,,c3,above_margin,above_margin,phi4_mini
288
+ donor_workload_lookup,math,gsm8k,,c4,memory_failure,memory_failure,phi4_mini
289
+ donor_workload_lookup,math,gsm8k,,c5,memory_failure,memory_failure,phi4_mini
290
+ donor_workload_lookup,code_heavy_tail,code_standard,,c0,within_margin,within_margin,qwen25_3b
291
+ donor_workload_lookup,code_heavy_tail,code_standard,,c1,within_margin,within_margin,qwen25_3b
292
+ donor_workload_lookup,code_heavy_tail,code_standard,,c2,within_margin,within_margin,qwen25_3b
293
+ donor_workload_lookup,code_heavy_tail,code_standard,,c3,within_margin,within_margin,qwen25_3b
294
+ donor_workload_lookup,code_heavy_tail,code_standard,,c4,memory_failure,memory_failure,qwen25_3b
295
+ donor_workload_lookup,code_heavy_tail,code_standard,,c5,memory_failure,memory_failure,qwen25_3b
296
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c0,within_margin,within_margin,qwen25_3b
297
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c1,within_margin,within_margin,qwen25_3b
298
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c2,within_margin,within_margin,qwen25_3b
299
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c3,within_margin,within_margin,qwen25_3b
300
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c4,memory_failure,memory_failure,qwen25_3b
301
+ donor_workload_lookup,code_heavy_tail,gsm8k,,c5,memory_failure,memory_failure,qwen25_3b
302
+ donor_workload_lookup,code_heavy_tail,math,,c0,within_margin,within_margin,qwen25_3b
303
+ donor_workload_lookup,code_heavy_tail,math,,c1,within_margin,within_margin,qwen25_3b
304
+ donor_workload_lookup,code_heavy_tail,math,,c2,within_margin,within_margin,qwen25_3b
305
+ donor_workload_lookup,code_heavy_tail,math,,c3,within_margin,within_margin,qwen25_3b
306
+ donor_workload_lookup,code_heavy_tail,math,,c4,memory_failure,memory_failure,qwen25_3b
307
+ donor_workload_lookup,code_heavy_tail,math,,c5,memory_failure,memory_failure,qwen25_3b
308
+ donor_workload_lookup,code_standard,code_heavy_tail,,c0,within_margin,within_margin,qwen25_3b
309
+ donor_workload_lookup,code_standard,code_heavy_tail,,c1,within_margin,within_margin,qwen25_3b
310
+ donor_workload_lookup,code_standard,code_heavy_tail,,c2,within_margin,within_margin,qwen25_3b
311
+ donor_workload_lookup,code_standard,code_heavy_tail,,c3,within_margin,within_margin,qwen25_3b
312
+ donor_workload_lookup,code_standard,code_heavy_tail,,c4,memory_failure,memory_failure,qwen25_3b
313
+ donor_workload_lookup,code_standard,code_heavy_tail,,c5,memory_failure,memory_failure,qwen25_3b
314
+ donor_workload_lookup,code_standard,gsm8k,,c0,within_margin,within_margin,qwen25_3b
315
+ donor_workload_lookup,code_standard,gsm8k,,c1,within_margin,within_margin,qwen25_3b
316
+ donor_workload_lookup,code_standard,gsm8k,,c2,within_margin,within_margin,qwen25_3b
317
+ donor_workload_lookup,code_standard,gsm8k,,c3,within_margin,within_margin,qwen25_3b
318
+ donor_workload_lookup,code_standard,gsm8k,,c4,memory_failure,memory_failure,qwen25_3b
319
+ donor_workload_lookup,code_standard,gsm8k,,c5,memory_failure,memory_failure,qwen25_3b
320
+ donor_workload_lookup,code_standard,math,,c0,within_margin,within_margin,qwen25_3b
321
+ donor_workload_lookup,code_standard,math,,c1,within_margin,within_margin,qwen25_3b
322
+ donor_workload_lookup,code_standard,math,,c2,within_margin,within_margin,qwen25_3b
323
+ donor_workload_lookup,code_standard,math,,c3,within_margin,within_margin,qwen25_3b
324
+ donor_workload_lookup,code_standard,math,,c4,memory_failure,memory_failure,qwen25_3b
325
+ donor_workload_lookup,code_standard,math,,c5,memory_failure,memory_failure,qwen25_3b
326
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c0,within_margin,within_margin,qwen25_3b
327
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c1,within_margin,within_margin,qwen25_3b
328
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c2,within_margin,within_margin,qwen25_3b
329
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c3,within_margin,within_margin,qwen25_3b
330
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c4,memory_failure,memory_failure,qwen25_3b
331
+ donor_workload_lookup,gsm8k,code_heavy_tail,,c5,memory_failure,memory_failure,qwen25_3b
332
+ donor_workload_lookup,gsm8k,code_standard,,c0,within_margin,within_margin,qwen25_3b
333
+ donor_workload_lookup,gsm8k,code_standard,,c1,within_margin,within_margin,qwen25_3b
334
+ donor_workload_lookup,gsm8k,code_standard,,c2,within_margin,within_margin,qwen25_3b
335
+ donor_workload_lookup,gsm8k,code_standard,,c3,within_margin,within_margin,qwen25_3b
336
+ donor_workload_lookup,gsm8k,code_standard,,c4,memory_failure,memory_failure,qwen25_3b
337
+ donor_workload_lookup,gsm8k,code_standard,,c5,memory_failure,memory_failure,qwen25_3b
338
+ donor_workload_lookup,gsm8k,math,,c0,within_margin,within_margin,qwen25_3b
339
+ donor_workload_lookup,gsm8k,math,,c1,within_margin,within_margin,qwen25_3b
340
+ donor_workload_lookup,gsm8k,math,,c2,within_margin,within_margin,qwen25_3b
341
+ donor_workload_lookup,gsm8k,math,,c3,within_margin,within_margin,qwen25_3b
342
+ donor_workload_lookup,gsm8k,math,,c4,memory_failure,memory_failure,qwen25_3b
343
+ donor_workload_lookup,gsm8k,math,,c5,memory_failure,memory_failure,qwen25_3b
344
+ donor_workload_lookup,math,code_heavy_tail,,c0,within_margin,within_margin,qwen25_3b
345
+ donor_workload_lookup,math,code_heavy_tail,,c1,within_margin,within_margin,qwen25_3b
346
+ donor_workload_lookup,math,code_heavy_tail,,c2,within_margin,within_margin,qwen25_3b
347
+ donor_workload_lookup,math,code_heavy_tail,,c3,within_margin,within_margin,qwen25_3b
348
+ donor_workload_lookup,math,code_heavy_tail,,c4,memory_failure,memory_failure,qwen25_3b
349
+ donor_workload_lookup,math,code_heavy_tail,,c5,memory_failure,memory_failure,qwen25_3b
350
+ donor_workload_lookup,math,code_standard,,c0,within_margin,within_margin,qwen25_3b
351
+ donor_workload_lookup,math,code_standard,,c1,within_margin,within_margin,qwen25_3b
352
+ donor_workload_lookup,math,code_standard,,c2,within_margin,within_margin,qwen25_3b
353
+ donor_workload_lookup,math,code_standard,,c3,within_margin,within_margin,qwen25_3b
354
+ donor_workload_lookup,math,code_standard,,c4,memory_failure,memory_failure,qwen25_3b
355
+ donor_workload_lookup,math,code_standard,,c5,memory_failure,memory_failure,qwen25_3b
356
+ donor_workload_lookup,math,gsm8k,,c0,within_margin,within_margin,qwen25_3b
357
+ donor_workload_lookup,math,gsm8k,,c1,within_margin,within_margin,qwen25_3b
358
+ donor_workload_lookup,math,gsm8k,,c2,within_margin,within_margin,qwen25_3b
359
+ donor_workload_lookup,math,gsm8k,,c3,within_margin,within_margin,qwen25_3b
360
+ donor_workload_lookup,math,gsm8k,,c4,memory_failure,memory_failure,qwen25_3b
361
+ donor_workload_lookup,math,gsm8k,,c5,memory_failure,memory_failure,qwen25_3b
362
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c2,within_margin,within_margin,granite33_2b
363
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c3,within_margin,within_margin,granite33_2b
364
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c4,above_margin,within_margin,granite33_2b
365
+ gpu_count_transfer,two_gpu,four_gpu,math,c2,within_margin,within_margin,granite33_2b
366
+ gpu_count_transfer,two_gpu,four_gpu,math,c3,within_margin,within_margin,granite33_2b
367
+ gpu_count_transfer,two_gpu,four_gpu,math,c4,above_margin,within_margin,granite33_2b
368
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c2,within_margin,within_margin,phi4_mini
369
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c3,above_margin,within_margin,phi4_mini
370
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c4,memory_failure,above_margin,phi4_mini
371
+ gpu_count_transfer,two_gpu,four_gpu,math,c2,within_margin,within_margin,phi4_mini
372
+ gpu_count_transfer,two_gpu,four_gpu,math,c3,above_margin,within_margin,phi4_mini
373
+ gpu_count_transfer,two_gpu,four_gpu,math,c4,memory_failure,above_margin,phi4_mini
374
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c2,within_margin,within_margin,qwen25_3b
375
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c3,within_margin,within_margin,qwen25_3b
376
+ gpu_count_transfer,two_gpu,four_gpu,gsm8k,c4,memory_failure,within_margin,qwen25_3b
377
+ gpu_count_transfer,two_gpu,four_gpu,math,c2,within_margin,within_margin,qwen25_3b
378
+ gpu_count_transfer,two_gpu,four_gpu,math,c3,within_margin,within_margin,qwen25_3b
379
+ gpu_count_transfer,two_gpu,four_gpu,math,c4,memory_failure,within_margin,qwen25_3b
380
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c2,within_margin,within_margin,granite33_2b
381
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c3,within_margin,within_margin,granite33_2b
382
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c4,within_margin,above_margin,granite33_2b
383
+ gpu_count_transfer,four_gpu,two_gpu,math,c2,within_margin,within_margin,granite33_2b
384
+ gpu_count_transfer,four_gpu,two_gpu,math,c3,within_margin,within_margin,granite33_2b
385
+ gpu_count_transfer,four_gpu,two_gpu,math,c4,within_margin,above_margin,granite33_2b
386
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c2,within_margin,within_margin,phi4_mini
387
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c3,within_margin,above_margin,phi4_mini
388
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c4,above_margin,memory_failure,phi4_mini
389
+ gpu_count_transfer,four_gpu,two_gpu,math,c2,within_margin,within_margin,phi4_mini
390
+ gpu_count_transfer,four_gpu,two_gpu,math,c3,within_margin,above_margin,phi4_mini
391
+ gpu_count_transfer,four_gpu,two_gpu,math,c4,above_margin,memory_failure,phi4_mini
392
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c2,within_margin,within_margin,qwen25_3b
393
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c3,within_margin,within_margin,qwen25_3b
394
+ gpu_count_transfer,four_gpu,two_gpu,gsm8k,c4,within_margin,memory_failure,qwen25_3b
395
+ gpu_count_transfer,four_gpu,two_gpu,math,c2,within_margin,within_margin,qwen25_3b
396
+ gpu_count_transfer,four_gpu,two_gpu,math,c3,within_margin,within_margin,qwen25_3b
397
+ gpu_count_transfer,four_gpu,two_gpu,math,c4,within_margin,memory_failure,qwen25_3b
tables/review_revision/decision-summaries.csv ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ rule,donor,target,decisions,correct_three_state,approved,approved_memory_failure,approved_above_margin,rejected_within_margin,target_within_margin,target_above_margin,target_memory_failure,within_margin_to_within_margin,within_margin_to_above_margin,within_margin_to_memory_failure,above_margin_to_within_margin,above_margin_to_above_margin,above_margin_to_memory_failure,memory_failure_to_within_margin,memory_failure_to_above_margin,memory_failure_to_memory_failure,model_family,included_donor_processes
2
+ source_model,granite33_2b,phi4_mini,24,16,16,0,4,0,12,4,8,12,4,0,0,0,4,0,0,4,,
3
+ source_model,granite33_2b,qwen25_3b,24,20,16,0,0,0,16,0,8,16,0,0,0,0,4,0,0,4,,
4
+ source_model,phi4_mini,granite33_2b,24,16,12,0,0,4,16,4,4,12,0,0,4,0,0,0,4,4,,
5
+ source_model,phi4_mini,qwen25_3b,24,20,12,0,0,4,16,0,8,12,0,0,4,0,0,0,0,8,,
6
+ source_model,qwen25_3b,granite33_2b,24,20,16,0,0,0,16,4,4,16,0,0,0,0,0,0,4,4,,
7
+ source_model,qwen25_3b,phi4_mini,24,20,16,0,4,0,12,4,8,12,4,0,0,0,0,0,0,8,,
8
+ donor_workload_lookup,code_heavy_tail,other_three_workloads,18,18,12,0,0,0,12,3,3,12,0,0,0,3,0,0,0,3,granite33_2b,18
9
+ donor_workload_lookup,code_standard,other_three_workloads,18,18,12,0,0,0,12,3,3,12,0,0,0,3,0,0,0,3,granite33_2b,18
10
+ donor_workload_lookup,gsm8k,other_three_workloads,18,18,12,0,0,0,12,3,3,12,0,0,0,3,0,0,0,3,granite33_2b,18
11
+ donor_workload_lookup,math,other_three_workloads,18,18,12,0,0,0,12,3,3,12,0,0,0,3,0,0,0,3,granite33_2b,18
12
+ donor_workload_lookup,code_heavy_tail,other_three_workloads,18,18,9,0,0,0,9,3,6,9,0,0,0,3,0,0,0,6,phi4_mini,18
13
+ donor_workload_lookup,code_standard,other_three_workloads,18,18,9,0,0,0,9,3,6,9,0,0,0,3,0,0,0,6,phi4_mini,18
14
+ donor_workload_lookup,gsm8k,other_three_workloads,18,18,9,0,0,0,9,3,6,9,0,0,0,3,0,0,0,6,phi4_mini,18
15
+ donor_workload_lookup,math,other_three_workloads,18,18,9,0,0,0,9,3,6,9,0,0,0,3,0,0,0,6,phi4_mini,18
16
+ donor_workload_lookup,code_heavy_tail,other_three_workloads,18,18,12,0,0,0,12,0,6,12,0,0,0,0,0,0,0,6,qwen25_3b,18
17
+ donor_workload_lookup,code_standard,other_three_workloads,18,18,12,0,0,0,12,0,6,12,0,0,0,0,0,0,0,6,qwen25_3b,18
18
+ donor_workload_lookup,gsm8k,other_three_workloads,18,18,12,0,0,0,12,0,6,12,0,0,0,0,0,0,0,6,qwen25_3b,18
19
+ donor_workload_lookup,math,other_three_workloads,18,18,12,0,0,0,12,0,6,12,0,0,0,0,0,0,0,6,qwen25_3b,18
20
+ gpu_count_transfer,two_gpu,four_gpu,18,10,10,0,0,6,16,2,0,10,0,0,4,0,0,2,2,0,,
21
+ gpu_count_transfer,four_gpu,two_gpu,18,10,16,2,4,0,10,4,4,10,4,2,0,0,2,0,0,0,,
tables/review_revision/factorial-allocation-history.csv ADDED
@@ -0,0 +1,145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_id,model_family,dataset,training_seed,actor_micro_batch,vllm_gpu_memory_utilization,job_id,hostname,captured_at_utc,source_head,working_copy_dirty
2
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s94,granite33_2b,code_heavy_tail,94,8,0.6,5031336,acn22.karolina.it4i.cz,2026-09-10T15:19:30.016434+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
3
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s95,granite33_2b,code_heavy_tail,95,8,0.6,5031337,acn22.karolina.it4i.cz,2026-09-10T15:19:31.029221+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
4
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s96,granite33_2b,code_heavy_tail,96,8,0.6,5031338,acn22.karolina.it4i.cz,2026-09-10T15:19:32.415732+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
5
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s94,granite33_2b,code_heavy_tail,94,16,0.6,5031339,acn22.karolina.it4i.cz,2026-09-10T15:19:34.146523+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
6
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s95,granite33_2b,code_heavy_tail,95,16,0.6,5031340,acn23.karolina.it4i.cz,2026-09-10T15:19:30.035186+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
7
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s96,granite33_2b,code_heavy_tail,96,16,0.6,5031341,acn23.karolina.it4i.cz,2026-09-10T15:19:30.884672+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
8
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s94,granite33_2b,code_heavy_tail,94,8,0.7,5031342,acn23.karolina.it4i.cz,2026-09-10T15:19:32.306349+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
9
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s95,granite33_2b,code_heavy_tail,95,8,0.7,5031327,acn23.karolina.it4i.cz,2026-09-10T15:19:34.082780+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
10
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s96,granite33_2b,code_heavy_tail,96,8,0.7,5031602,acn34.karolina.it4i.cz,2026-09-10T15:49:00.545807+00:00,beee710417101dc2450688b5310fd5eab08172f1,1
11
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s94,granite33_2b,code_heavy_tail,94,16,0.7,5031377,acn26.karolina.it4i.cz,2026-09-10T15:24:37.979256+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
12
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s95,granite33_2b,code_heavy_tail,95,16,0.7,5031590,acn35.karolina.it4i.cz,2026-09-10T15:49:01.912210+00:00,beee710417101dc2450688b5310fd5eab08172f1,1
13
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s96,granite33_2b,code_heavy_tail,96,16,0.7,5031379,acn26.karolina.it4i.cz,2026-09-10T15:24:41.230075+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
14
+ strong-factorial-granite33_2b-grpo-code_standard-base-s94,granite33_2b,code_standard,94,8,0.6,5031380,acn21.karolina.it4i.cz,2026-09-10T15:24:37.066167+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
15
+ strong-factorial-granite33_2b-grpo-code_standard-base-s95,granite33_2b,code_standard,95,8,0.6,5031372,acn21.karolina.it4i.cz,2026-09-10T15:24:38.018701+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
16
+ strong-factorial-granite33_2b-grpo-code_standard-base-s96,granite33_2b,code_standard,96,8,0.6,5031316,acn38.karolina.it4i.cz,2026-09-10T15:18:58.250392+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
17
+ strong-factorial-granite33_2b-grpo-code_standard-actor_only-s94,granite33_2b,code_standard,94,16,0.6,5031317,acn38.karolina.it4i.cz,2026-09-10T15:18:59.435966+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
18
+ strong-factorial-granite33_2b-grpo-code_standard-actor_only-s95,granite33_2b,code_standard,95,16,0.6,5031318,acn38.karolina.it4i.cz,2026-09-10T15:19:01.434356+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
19
+ strong-factorial-granite33_2b-grpo-code_standard-actor_only-s96,granite33_2b,code_standard,96,16,0.6,5031319,acn32.karolina.it4i.cz,2026-09-10T15:18:59.253033+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
20
+ strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s94,granite33_2b,code_standard,94,8,0.7,5031320,acn32.karolina.it4i.cz,2026-09-10T15:19:00.090494+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
21
+ strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s95,granite33_2b,code_standard,95,8,0.7,5031321,acn32.karolina.it4i.cz,2026-09-10T15:19:01.540758+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
22
+ strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s96,granite33_2b,code_standard,96,8,0.7,5031322,acn32.karolina.it4i.cz,2026-09-10T15:19:03.460426+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
23
+ strong-factorial-granite33_2b-grpo-code_standard-joint-s94,granite33_2b,code_standard,94,16,0.7,5031323,acn33.karolina.it4i.cz,2026-09-10T15:18:58.287965+00:00,30f1ad25f6facf7c8d2e89d5fe882109dfda1207,1
24
+ strong-factorial-granite33_2b-grpo-code_standard-joint-s95,granite33_2b,code_standard,95,16,0.7,5031351,acn38.karolina.it4i.cz,2026-09-10T15:22:34.412817+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
25
+ strong-factorial-granite33_2b-grpo-code_standard-joint-s96,granite33_2b,code_standard,96,16,0.7,5031352,acn38.karolina.it4i.cz,2026-09-10T15:22:35.920890+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
26
+ strong-factorial-granite33_2b-grpo-gsm8k_long-base-s94,granite33_2b,gsm8k,94,8,0.6,5031353,acn32.karolina.it4i.cz,2026-09-10T15:22:34.240546+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
27
+ strong-factorial-granite33_2b-grpo-gsm8k_long-base-s95,granite33_2b,gsm8k,95,8,0.6,5031354,acn32.karolina.it4i.cz,2026-09-10T15:22:35.096397+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
28
+ strong-factorial-granite33_2b-grpo-gsm8k_long-base-s96,granite33_2b,gsm8k,96,8,0.6,5031355,acn32.karolina.it4i.cz,2026-09-10T15:22:36.550290+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
29
+ strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s94,granite33_2b,gsm8k,94,16,0.6,5031356,acn32.karolina.it4i.cz,2026-09-10T15:22:38.350806+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
30
+ strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s95,granite33_2b,gsm8k,95,16,0.6,5031357,acn33.karolina.it4i.cz,2026-09-10T15:22:33.444704+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
31
+ strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s96,granite33_2b,gsm8k,96,16,0.6,5031366,acn38.karolina.it4i.cz,2026-09-10T15:23:04.877482+00:00,f2efa5f89d5d773a81e7a2d34a12939c9435bf87,1
32
+ strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s94,granite33_2b,gsm8k,94,8,0.7,5031393,acn38.karolina.it4i.cz,2026-09-10T15:26:09.157534+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
33
+ strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s95,granite33_2b,gsm8k,95,8,0.7,5031394,acn38.karolina.it4i.cz,2026-09-10T15:26:10.618009+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
34
+ strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s96,granite33_2b,gsm8k,96,8,0.7,5031400,acn38.karolina.it4i.cz,2026-09-10T15:26:41.277828+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
35
+ strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s94,granite33_2b,gsm8k,94,16,0.7,5031401,acn32.karolina.it4i.cz,2026-09-10T15:26:40.849884+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
36
+ strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s95,granite33_2b,gsm8k,95,16,0.7,5031402,acn32.karolina.it4i.cz,2026-09-10T15:26:39.571229+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
37
+ strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s96,granite33_2b,gsm8k,96,16,0.7,5031403,acn33.karolina.it4i.cz,2026-09-10T15:26:39.518860+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
38
+ strong-factorial-granite33_2b-grpo-math_short-base-s94,granite33_2b,math,94,8,0.6,5031406,acn32.karolina.it4i.cz,2026-09-10T15:27:10.571406+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
39
+ strong-factorial-granite33_2b-grpo-math_short-base-s95,granite33_2b,math,95,8,0.6,5031407,acn32.karolina.it4i.cz,2026-09-10T15:27:11.826894+00:00,ae8c9de45affc66f0a1d1c31e6836e9538e2986c,1
40
+ strong-factorial-granite33_2b-grpo-math_short-base-s96,granite33_2b,math,96,8,0.6,5031428,acn28.karolina.it4i.cz,2026-09-10T15:29:43.652356+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
41
+ strong-factorial-granite33_2b-grpo-math_short-actor_only-s94,granite33_2b,math,94,16,0.6,5031429,acn38.karolina.it4i.cz,2026-09-10T15:30:14.265881+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
42
+ strong-factorial-granite33_2b-grpo-math_short-actor_only-s95,granite33_2b,math,95,16,0.6,5031430,acn38.karolina.it4i.cz,2026-09-10T15:30:15.364541+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
43
+ strong-factorial-granite33_2b-grpo-math_short-actor_only-s96,granite33_2b,math,96,16,0.6,5031431,acn38.karolina.it4i.cz,2026-09-10T15:30:16.889441+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
44
+ strong-factorial-granite33_2b-grpo-math_short-reservation_only-s94,granite33_2b,math,94,8,0.7,5031432,acn22.karolina.it4i.cz,2026-09-10T15:30:15.188903+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
45
+ strong-factorial-granite33_2b-grpo-math_short-reservation_only-s95,granite33_2b,math,95,8,0.7,5031435,acn22.karolina.it4i.cz,2026-09-10T15:30:45.988919+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
46
+ strong-factorial-granite33_2b-grpo-math_short-reservation_only-s96,granite33_2b,math,96,8,0.7,5031438,acn22.karolina.it4i.cz,2026-09-10T15:31:17.567979+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
47
+ strong-factorial-granite33_2b-grpo-math_short-joint-s94,granite33_2b,math,94,16,0.7,5031439,acn22.karolina.it4i.cz,2026-09-10T15:31:18.984339+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
48
+ strong-factorial-granite33_2b-grpo-math_short-joint-s95,granite33_2b,math,95,16,0.7,5031497,acn38.karolina.it4i.cz,2026-09-10T15:33:51.334156+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
49
+ strong-factorial-granite33_2b-grpo-math_short-joint-s96,granite33_2b,math,96,16,0.7,5031498,acn23.karolina.it4i.cz,2026-09-10T15:33:49.808234+00:00,907f8f908cde54bc18374f93e5972d6e5059387e,1
50
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s94,phi4_mini,code_heavy_tail,94,8,0.6,5031694,acn05.karolina.it4i.cz,2026-09-10T15:59:13.232315+00:00,0208d66171970a7fed998ab73fab4bf1a54f423a,1
51
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s95,phi4_mini,code_heavy_tail,95,8,0.6,5031695,acn05.karolina.it4i.cz,2026-09-10T15:59:10.821012+00:00,0208d66171970a7fed998ab73fab4bf1a54f423a,1
52
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s96,phi4_mini,code_heavy_tail,96,8,0.6,5031700,acn03.karolina.it4i.cz,2026-09-10T15:59:40.110233+00:00,0208d66171970a7fed998ab73fab4bf1a54f423a,1
53
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s94,phi4_mini,code_heavy_tail,94,16,0.6,5031706,acn06.karolina.it4i.cz,2026-09-10T16:00:10.904209+00:00,0208d66171970a7fed998ab73fab4bf1a54f423a,1
54
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s95,phi4_mini,code_heavy_tail,95,16,0.6,5031718,acn15.karolina.it4i.cz,2026-09-10T16:02:13.393744+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
55
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s96,phi4_mini,code_heavy_tail,96,16,0.6,5031719,acn15.karolina.it4i.cz,2026-09-10T16:02:12.015726+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
56
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s94,phi4_mini,code_heavy_tail,94,8,0.7,5031720,acn15.karolina.it4i.cz,2026-09-10T16:02:14.763522+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
57
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s95,phi4_mini,code_heavy_tail,95,8,0.7,5031721,acn19.karolina.it4i.cz,2026-09-10T16:02:13.630505+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
58
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s96,phi4_mini,code_heavy_tail,96,8,0.7,5031728,acn03.karolina.it4i.cz,2026-09-10T16:03:43.683802+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
59
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s94,phi4_mini,code_heavy_tail,94,16,0.7,5031729,acn19.karolina.it4i.cz,2026-09-10T16:03:43.312451+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
60
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s95,phi4_mini,code_heavy_tail,95,16,0.7,5031730,acn22.karolina.it4i.cz,2026-09-10T16:03:43.370142+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
61
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s96,phi4_mini,code_heavy_tail,96,16,0.7,5031737,acn06.karolina.it4i.cz,2026-09-10T16:04:13.588554+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
62
+ strong-factorial-phi4_mini-grpo-code_standard-base-s94,phi4_mini,code_standard,94,8,0.6,5031745,acn06.karolina.it4i.cz,2026-09-10T16:06:15.635233+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
63
+ strong-factorial-phi4_mini-grpo-code_standard-base-s95,phi4_mini,code_standard,95,8,0.6,5031746,acn15.karolina.it4i.cz,2026-09-10T16:06:15.009989+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
64
+ strong-factorial-phi4_mini-grpo-code_standard-base-s96,phi4_mini,code_standard,96,8,0.6,5031747,acn15.karolina.it4i.cz,2026-09-10T16:06:16.188754+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
65
+ strong-factorial-phi4_mini-grpo-code_standard-actor_only-s94,phi4_mini,code_standard,94,16,0.6,5031749,acn03.karolina.it4i.cz,2026-09-10T16:06:45.750195+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
66
+ strong-factorial-phi4_mini-grpo-code_standard-actor_only-s95,phi4_mini,code_standard,95,16,0.6,5031754,acn03.karolina.it4i.cz,2026-09-10T16:07:46.403198+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
67
+ strong-factorial-phi4_mini-grpo-code_standard-actor_only-s96,phi4_mini,code_standard,96,16,0.6,5031755,acn06.karolina.it4i.cz,2026-09-10T16:07:46.336071+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
68
+ strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s94,phi4_mini,code_standard,94,8,0.7,5031756,acn19.karolina.it4i.cz,2026-09-10T16:07:46.014668+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
69
+ strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s95,phi4_mini,code_standard,95,8,0.7,5031759,acn06.karolina.it4i.cz,2026-09-10T16:08:16.667368+00:00,117c94f69a85d95955a140e4662c669c2176c8eb,1
70
+ strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s96,phi4_mini,code_standard,96,8,0.7,5031767,acn06.karolina.it4i.cz,2026-09-10T16:10:18.033201+00:00,e4b3b633567c1393f498b83cff0fdc3b9ce8b04b,1
71
+ strong-factorial-phi4_mini-grpo-code_standard-joint-s94,phi4_mini,code_standard,94,16,0.7,5031768,acn15.karolina.it4i.cz,2026-09-10T16:10:20.380707+00:00,e4b3b633567c1393f498b83cff0fdc3b9ce8b04b,1
72
+ strong-factorial-phi4_mini-grpo-code_standard-joint-s95,phi4_mini,code_standard,95,16,0.7,5031769,acn15.karolina.it4i.cz,2026-09-10T16:10:17.951820+00:00,e4b3b633567c1393f498b83cff0fdc3b9ce8b04b,1
73
+ strong-factorial-phi4_mini-grpo-code_standard-joint-s96,phi4_mini,code_standard,96,16,0.7,5031772,acn03.karolina.it4i.cz,2026-09-10T16:10:48.372944+00:00,e4b3b633567c1393f498b83cff0fdc3b9ce8b04b,1
74
+ strong-factorial-phi4_mini-grpo-gsm8k-base-s94,phi4_mini,gsm8k,94,8,0.6,5031778,acn03.karolina.it4i.cz,2026-09-10T16:11:49.822374+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
75
+ strong-factorial-phi4_mini-grpo-gsm8k-base-s95,phi4_mini,gsm8k,95,8,0.6,5031779,acn06.karolina.it4i.cz,2026-09-10T16:11:49.158987+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
76
+ strong-factorial-phi4_mini-grpo-gsm8k-base-s96,phi4_mini,gsm8k,96,8,0.6,5031780,acn15.karolina.it4i.cz,2026-09-10T16:11:49.190404+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
77
+ strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s94,phi4_mini,gsm8k,94,16,0.6,5031783,acn04.karolina.it4i.cz,2026-09-10T16:12:20.200025+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
78
+ strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s95,phi4_mini,gsm8k,95,16,0.6,5031791,acn04.karolina.it4i.cz,2026-09-10T16:13:50.500858+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
79
+ strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s96,phi4_mini,gsm8k,96,16,0.6,5031796,acn06.karolina.it4i.cz,2026-09-10T16:14:20.724694+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
80
+ strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s94,phi4_mini,gsm8k,94,8,0.7,5031797,acn06.karolina.it4i.cz,2026-09-10T16:14:22.039877+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
81
+ strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s95,phi4_mini,gsm8k,95,8,0.7,5031801,acn09.karolina.it4i.cz,2026-09-10T16:14:53.056268+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,0
82
+ strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s96,phi4_mini,gsm8k,96,8,0.7,5031806,acn06.karolina.it4i.cz,2026-09-10T16:15:52.008929+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,1
83
+ strong-factorial-phi4_mini-grpo-gsm8k-joint-s94,phi4_mini,gsm8k,94,16,0.7,5031807,acn28.karolina.it4i.cz,2026-09-10T16:15:52.127512+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,1
84
+ strong-factorial-phi4_mini-grpo-gsm8k-joint-s95,phi4_mini,gsm8k,95,16,0.7,5031808,acn28.karolina.it4i.cz,2026-09-10T16:15:53.447168+00:00,104e6a4d302a12300c14f7dfe96291427179e7e5,1
85
+ strong-factorial-phi4_mini-grpo-gsm8k-joint-s96,phi4_mini,gsm8k,96,16,0.7,5031812,acn04.karolina.it4i.cz,2026-09-10T16:16:23.568508+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,0
86
+ strong-factorial-phi4_mini-grpo-math-base-s94,phi4_mini,math,94,8,0.6,5031823,acn04.karolina.it4i.cz,2026-09-10T16:17:53.542160+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,0
87
+ strong-factorial-phi4_mini-grpo-math-base-s95,phi4_mini,math,95,8,0.6,5031824,acn06.karolina.it4i.cz,2026-09-10T16:17:53.552290+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,0
88
+ strong-factorial-phi4_mini-grpo-math-base-s96,phi4_mini,math,96,8,0.6,5031826,acn04.karolina.it4i.cz,2026-09-10T16:18:23.476981+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,0
89
+ strong-factorial-phi4_mini-grpo-math-actor_only-s94,phi4_mini,math,94,16,0.6,5031829,acn19.karolina.it4i.cz,2026-09-10T16:18:54.009782+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,0
90
+ strong-factorial-phi4_mini-grpo-math-actor_only-s95,phi4_mini,math,95,16,0.6,5031836,acn04.karolina.it4i.cz,2026-09-10T16:19:54.791741+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
91
+ strong-factorial-phi4_mini-grpo-math-actor_only-s96,phi4_mini,math,96,16,0.6,5031837,acn06.karolina.it4i.cz,2026-09-10T16:19:54.712548+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
92
+ strong-factorial-phi4_mini-grpo-math-reservation_only-s94,phi4_mini,math,94,8,0.7,5031838,acn19.karolina.it4i.cz,2026-09-10T16:19:55.870109+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
93
+ strong-factorial-phi4_mini-grpo-math-reservation_only-s95,phi4_mini,math,95,8,0.7,5031841,acn04.karolina.it4i.cz,2026-09-10T16:20:25.697705+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
94
+ strong-factorial-phi4_mini-grpo-math-reservation_only-s96,phi4_mini,math,96,8,0.7,5031856,acn04.karolina.it4i.cz,2026-09-10T16:21:56.158026+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
95
+ strong-factorial-phi4_mini-grpo-math-joint-s94,phi4_mini,math,94,16,0.7,5031857,acn06.karolina.it4i.cz,2026-09-10T16:21:56.203942+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
96
+ strong-factorial-phi4_mini-grpo-math-joint-s95,phi4_mini,math,95,16,0.7,5031860,acn04.karolina.it4i.cz,2026-09-10T16:22:26.186960+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
97
+ strong-factorial-phi4_mini-grpo-math-joint-s96,phi4_mini,math,96,16,0.7,5031862,acn03.karolina.it4i.cz,2026-09-10T16:22:56.654839+00:00,1a60a4e2469617d2d20ad2cf087d7331d3ca30fd,1
98
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s94,qwen25_3b,code_heavy_tail,94,8,0.6,5031981,acn28.karolina.it4i.cz,2026-09-10T16:45:13.873196+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,0
99
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s95,qwen25_3b,code_heavy_tail,95,8,0.6,5031982,acn40.karolina.it4i.cz,2026-09-10T16:45:12.976324+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,0
100
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s96,qwen25_3b,code_heavy_tail,96,8,0.6,5031983,acn40.karolina.it4i.cz,2026-09-10T16:45:14.530015+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,0
101
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s94,qwen25_3b,code_heavy_tail,94,16,0.6,5031985,acn40.karolina.it4i.cz,2026-09-10T16:45:44.279676+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,0
102
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s95,qwen25_3b,code_heavy_tail,95,16,0.6,5031987,acn28.karolina.it4i.cz,2026-09-10T16:47:45.109007+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,1
103
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s96,qwen25_3b,code_heavy_tail,96,16,0.6,5031988,acn28.karolina.it4i.cz,2026-09-10T16:47:46.256770+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,1
104
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s94,qwen25_3b,code_heavy_tail,94,8,0.7,5031989,acn28.karolina.it4i.cz,2026-09-10T16:48:16.011342+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,1
105
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s95,qwen25_3b,code_heavy_tail,95,8,0.7,5031990,acn37.karolina.it4i.cz,2026-09-10T16:48:15.511470+00:00,e423173334f8469735f1c6a5b757b3feb136ee19,1
106
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s96,qwen25_3b,code_heavy_tail,96,8,0.7,5031991,acn28.karolina.it4i.cz,2026-09-10T16:49:16.409169+00:00,bc4caf46b4196891c3da287974e94fa016a70cd8,0
107
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s94,qwen25_3b,code_heavy_tail,94,16,0.7,5031992,acn40.karolina.it4i.cz,2026-09-10T16:49:17.351467+00:00,bc4caf46b4196891c3da287974e94fa016a70cd8,0
108
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s95,qwen25_3b,code_heavy_tail,95,16,0.7,5031993,acn40.karolina.it4i.cz,2026-09-10T16:49:16.292127+00:00,bc4caf46b4196891c3da287974e94fa016a70cd8,0
109
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s96,qwen25_3b,code_heavy_tail,96,16,0.7,5031994,acn40.karolina.it4i.cz,2026-09-10T16:49:47.080579+00:00,bc4caf46b4196891c3da287974e94fa016a70cd8,0
110
+ strong-factorial-qwen25_3b-grpo-code_standard-base-s94,qwen25_3b,code_standard,94,8,0.6,5032005,acn03.karolina.it4i.cz,2026-09-10T16:51:48.188803+00:00,85a02245911d690fdd7f256b7bfbea0e5cf92656,0
111
+ strong-factorial-qwen25_3b-grpo-code_standard-base-s95,qwen25_3b,code_standard,95,8,0.6,5032006,acn03.karolina.it4i.cz,2026-09-10T16:51:49.368256+00:00,85a02245911d690fdd7f256b7bfbea0e5cf92656,0
112
+ strong-factorial-qwen25_3b-grpo-code_standard-base-s96,qwen25_3b,code_standard,96,8,0.6,5032007,acn28.karolina.it4i.cz,2026-09-10T16:51:48.257109+00:00,85a02245911d690fdd7f256b7bfbea0e5cf92656,0
113
+ strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s94,qwen25_3b,code_standard,94,16,0.6,5032012,acn06.karolina.it4i.cz,2026-09-10T16:52:18.564778+00:00,85a02245911d690fdd7f256b7bfbea0e5cf92656,1
114
+ strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s95,qwen25_3b,code_standard,95,16,0.6,5032015,acn07.karolina.it4i.cz,2026-09-10T16:53:19.416551+00:00,85a02245911d690fdd7f256b7bfbea0e5cf92656,1
115
+ strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s96,qwen25_3b,code_standard,96,16,0.6,5032016,acn07.karolina.it4i.cz,2026-09-10T16:53:20.480202+00:00,85a02245911d690fdd7f256b7bfbea0e5cf92656,1
116
+ strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s94,qwen25_3b,code_standard,94,8,0.7,5032033,acn38.karolina.it4i.cz,2026-09-10T16:58:54.462153+00:00,e5835941def3926967612e72a54bfaff3eb694c4,0
117
+ strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s95,qwen25_3b,code_standard,95,8,0.7,5032018,acn28.karolina.it4i.cz,2026-09-10T16:53:49.954140+00:00,85a02245911d690fdd7f256b7bfbea0e5cf92656,1
118
+ strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s96,qwen25_3b,code_standard,96,8,0.7,5032020,acn04.karolina.it4i.cz,2026-09-10T16:54:51.332339+00:00,1fa29f66785f16576a97775d481d9c545a30e042,0
119
+ strong-factorial-qwen25_3b-grpo-code_standard-joint-s94,qwen25_3b,code_standard,94,16,0.7,5032024,acn28.karolina.it4i.cz,2026-09-10T16:55:51.510033+00:00,1fa29f66785f16576a97775d481d9c545a30e042,0
120
+ strong-factorial-qwen25_3b-grpo-code_standard-joint-s95,qwen25_3b,code_standard,95,16,0.7,5032025,acn40.karolina.it4i.cz,2026-09-10T16:55:51.234901+00:00,1fa29f66785f16576a97775d481d9c545a30e042,0
121
+ strong-factorial-qwen25_3b-grpo-code_standard-joint-s96,qwen25_3b,code_standard,96,16,0.7,5032026,acn40.karolina.it4i.cz,2026-09-10T16:55:52.407725+00:00,1fa29f66785f16576a97775d481d9c545a30e042,0
122
+ strong-factorial-qwen25_3b-grpo-gsm8k-base-s94,qwen25_3b,gsm8k,94,8,0.6,5032028,acn40.karolina.it4i.cz,2026-09-10T16:56:22.857577+00:00,1fa29f66785f16576a97775d481d9c545a30e042,1
123
+ strong-factorial-qwen25_3b-grpo-gsm8k-base-s95,qwen25_3b,gsm8k,95,8,0.6,5032029,acn37.karolina.it4i.cz,2026-09-10T16:56:52.054687+00:00,26714705ca8c72e6cb4c8c9a5a229bcb9f081221,0
124
+ strong-factorial-qwen25_3b-grpo-gsm8k-base-s96,qwen25_3b,gsm8k,96,8,0.6,5032030,acn37.karolina.it4i.cz,2026-09-10T16:57:22.238847+00:00,26714705ca8c72e6cb4c8c9a5a229bcb9f081221,0
125
+ strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s94,qwen25_3b,gsm8k,94,16,0.6,5032031,acn28.karolina.it4i.cz,2026-09-10T16:57:53.194031+00:00,26714705ca8c72e6cb4c8c9a5a229bcb9f081221,0
126
+ strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s95,qwen25_3b,gsm8k,95,16,0.6,5032034,acn37.karolina.it4i.cz,2026-09-10T16:58:53.495563+00:00,e5835941def3926967612e72a54bfaff3eb694c4,0
127
+ strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s96,qwen25_3b,gsm8k,96,16,0.6,5032036,acn28.karolina.it4i.cz,2026-09-10T16:59:54.566443+00:00,e5835941def3926967612e72a54bfaff3eb694c4,0
128
+ strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s94,qwen25_3b,gsm8k,94,8,0.7,5032037,acn38.karolina.it4i.cz,2026-09-10T16:59:55.663410+00:00,e5835941def3926967612e72a54bfaff3eb694c4,0
129
+ strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s95,qwen25_3b,gsm8k,95,8,0.7,5032038,acn38.karolina.it4i.cz,2026-09-10T16:59:54.522276+00:00,e5835941def3926967612e72a54bfaff3eb694c4,0
130
+ strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s96,qwen25_3b,gsm8k,96,8,0.7,5032039,acn38.karolina.it4i.cz,2026-09-10T17:00:25.657498+00:00,e5835941def3926967612e72a54bfaff3eb694c4,1
131
+ strong-factorial-qwen25_3b-grpo-gsm8k-joint-s94,qwen25_3b,gsm8k,94,16,0.7,5032040,acn37.karolina.it4i.cz,2026-09-10T17:00:55.319472+00:00,e5835941def3926967612e72a54bfaff3eb694c4,1
132
+ strong-factorial-qwen25_3b-grpo-gsm8k-joint-s95,qwen25_3b,gsm8k,95,16,0.7,5032041,acn37.karolina.it4i.cz,2026-09-10T17:00:56.523261+00:00,e5835941def3926967612e72a54bfaff3eb694c4,1
133
+ strong-factorial-qwen25_3b-grpo-gsm8k-joint-s96,qwen25_3b,gsm8k,96,16,0.7,5032042,acn28.karolina.it4i.cz,2026-09-10T17:01:25.592399+00:00,e5835941def3926967612e72a54bfaff3eb694c4,1
134
+ strong-factorial-qwen25_3b-grpo-math-base-s94,qwen25_3b,math,94,8,0.6,5032043,acn37.karolina.it4i.cz,2026-09-10T17:02:26.374390+00:00,e5835941def3926967612e72a54bfaff3eb694c4,1
135
+ strong-factorial-qwen25_3b-grpo-math-base-s95,qwen25_3b,math,95,8,0.6,5032045,acn28.karolina.it4i.cz,2026-09-10T17:03:27.244174+00:00,b2c28c9ef3d52cb831536838fee4bb63661eae20,0
136
+ strong-factorial-qwen25_3b-grpo-math-base-s96,qwen25_3b,math,96,8,0.6,5032046,acn38.karolina.it4i.cz,2026-09-10T17:03:58.041347+00:00,b2c28c9ef3d52cb831536838fee4bb63661eae20,1
137
+ strong-factorial-qwen25_3b-grpo-math-actor_only-s94,qwen25_3b,math,94,16,0.6,5032047,acn38.karolina.it4i.cz,2026-09-10T17:04:00.347104+00:00,b2c28c9ef3d52cb831536838fee4bb63661eae20,1
138
+ strong-factorial-qwen25_3b-grpo-math-actor_only-s95,qwen25_3b,math,95,16,0.6,5032051,acn38.karolina.it4i.cz,2026-09-10T17:04:29.164374+00:00,cb73d3842d133cacae4a5b4dcfc5daf74b3f930f,0
139
+ strong-factorial-qwen25_3b-grpo-math-actor_only-s96,qwen25_3b,math,96,16,0.6,5032052,acn28.karolina.it4i.cz,2026-09-10T17:04:58.546383+00:00,d34c26357a12248b10920278eeca38baa6ec76e5,0
140
+ strong-factorial-qwen25_3b-grpo-math-reservation_only-s94,qwen25_3b,math,94,8,0.7,5032053,acn40.karolina.it4i.cz,2026-09-10T17:04:58.379386+00:00,d34c26357a12248b10920278eeca38baa6ec76e5,0
141
+ strong-factorial-qwen25_3b-grpo-math-reservation_only-s95,qwen25_3b,math,95,8,0.7,5032054,acn40.karolina.it4i.cz,2026-09-10T17:04:59.389955+00:00,d34c26357a12248b10920278eeca38baa6ec76e5,0
142
+ strong-factorial-qwen25_3b-grpo-math-reservation_only-s96,qwen25_3b,math,96,8,0.7,5032056,acn40.karolina.it4i.cz,2026-09-10T17:06:29.657335+00:00,d34c26357a12248b10920278eeca38baa6ec76e5,1
143
+ strong-factorial-qwen25_3b-grpo-math-joint-s94,qwen25_3b,math,94,16,0.7,5032058,acn28.karolina.it4i.cz,2026-09-10T17:07:30.551788+00:00,c836159538fb835d011c7c97c51fb2b26057a4ff,0
144
+ strong-factorial-qwen25_3b-grpo-math-joint-s95,qwen25_3b,math,95,16,0.7,5032059,acn37.karolina.it4i.cz,2026-09-10T17:08:00.662267+00:00,c836159538fb835d011c7c97c51fb2b26057a4ff,0
145
+ strong-factorial-qwen25_3b-grpo-math-joint-s96,qwen25_3b,math,96,16,0.7,5032060,acn37.karolina.it4i.cz,2026-09-10T17:08:01.735844+00:00,c836159538fb835d011c7c97c51fb2b26057a4ff,0
tables/review_revision/factorial-peak-stages.csv ADDED
@@ -0,0 +1,145 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_id,model_family,dataset,training_seed,actor_micro_batch,vllm_gpu_memory_utilization,success,job_id,environment_json,whole_run_peak_mib,peak_setting_stages,actor_sets_whole_run_peak
2
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s94,granite33_2b,code_heavy_tail,94,8,0.6,1,5031336,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s94/environment-5031336.json,32544.0,rollout,0
3
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s95,granite33_2b,code_heavy_tail,95,8,0.6,1,5031337,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s95/environment-5031337.json,32536.0,rollout,0
4
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s96,granite33_2b,code_heavy_tail,96,8,0.6,1,5031338,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-base-s96/environment-5031338.json,32744.0,rollout,0
5
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s94,granite33_2b,code_heavy_tail,94,16,0.6,1,5031339,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s94/environment-5031339.json,32534.0,rollout,0
6
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s95,granite33_2b,code_heavy_tail,95,16,0.6,1,5031340,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s95/environment-5031340.json,32542.0,rollout,0
7
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s96,granite33_2b,code_heavy_tail,96,16,0.6,1,5031341,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-actor_only-s96/environment-5031341.json,32536.0,rollout,0
8
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s94,granite33_2b,code_heavy_tail,94,8,0.7,1,5031342,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s94/environment-5031342.json,36572.0,rollout,0
9
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s95,granite33_2b,code_heavy_tail,95,8,0.7,1,5031327,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s95/environment-5031327.json,36534.0,rollout,0
10
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s96,granite33_2b,code_heavy_tail,96,8,0.7,1,5031602,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-reservation_only-s96/environment-5031602.json,36542.0,rollout,0
11
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s94,granite33_2b,code_heavy_tail,94,16,0.7,1,5031377,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s94/environment-5031377.json,36532.0,rollout,0
12
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s95,granite33_2b,code_heavy_tail,95,16,0.7,1,5031590,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s95/environment-5031590.json,36542.0,rollout,0
13
+ strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s96,granite33_2b,code_heavy_tail,96,16,0.7,1,5031379,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_heavy_tail-joint-s96/environment-5031379.json,36536.0,rollout,0
14
+ strong-factorial-granite33_2b-grpo-code_standard-base-s94,granite33_2b,code_standard,94,8,0.6,1,5031380,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-base-s94/environment-5031380.json,32506.0,rollout,0
15
+ strong-factorial-granite33_2b-grpo-code_standard-base-s95,granite33_2b,code_standard,95,8,0.6,1,5031372,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-base-s95/environment-5031372.json,32302.0,idle;weight_sync,0
16
+ strong-factorial-granite33_2b-grpo-code_standard-base-s96,granite33_2b,code_standard,96,8,0.6,1,5031316,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-base-s96/environment-5031316.json,32302.0,idle;weight_sync,0
17
+ strong-factorial-granite33_2b-grpo-code_standard-actor_only-s94,granite33_2b,code_standard,94,16,0.6,1,5031317,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-actor_only-s94/environment-5031317.json,32426.0,rollout,0
18
+ strong-factorial-granite33_2b-grpo-code_standard-actor_only-s95,granite33_2b,code_standard,95,16,0.6,1,5031318,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-actor_only-s95/environment-5031318.json,32534.0,rollout,0
19
+ strong-factorial-granite33_2b-grpo-code_standard-actor_only-s96,granite33_2b,code_standard,96,16,0.6,1,5031319,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-actor_only-s96/environment-5031319.json,32554.0,rollout,0
20
+ strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s94,granite33_2b,code_standard,94,8,0.7,1,5031320,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s94/environment-5031320.json,36506.0,rollout,0
21
+ strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s95,granite33_2b,code_standard,95,8,0.7,1,5031321,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s95/environment-5031321.json,36506.0,rollout,0
22
+ strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s96,granite33_2b,code_standard,96,8,0.7,1,5031322,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-reservation_only-s96/environment-5031322.json,36548.0,rollout,0
23
+ strong-factorial-granite33_2b-grpo-code_standard-joint-s94,granite33_2b,code_standard,94,16,0.7,1,5031323,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-joint-s94/environment-5031323.json,36322.0,rollout,0
24
+ strong-factorial-granite33_2b-grpo-code_standard-joint-s95,granite33_2b,code_standard,95,16,0.7,1,5031351,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-joint-s95/environment-5031351.json,36458.0,rollout,0
25
+ strong-factorial-granite33_2b-grpo-code_standard-joint-s96,granite33_2b,code_standard,96,16,0.7,1,5031352,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-code_standard-joint-s96/environment-5031352.json,36392.0,rollout,0
26
+ strong-factorial-granite33_2b-grpo-gsm8k_long-base-s94,granite33_2b,gsm8k,94,8,0.6,1,5031353,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-base-s94/environment-5031353.json,32322.0,idle,0
27
+ strong-factorial-granite33_2b-grpo-gsm8k_long-base-s95,granite33_2b,gsm8k,95,8,0.6,1,5031354,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-base-s95/environment-5031354.json,32322.0,idle;weight_sync,0
28
+ strong-factorial-granite33_2b-grpo-gsm8k_long-base-s96,granite33_2b,gsm8k,96,8,0.6,1,5031355,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-base-s96/environment-5031355.json,32322.0,idle;weight_sync,0
29
+ strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s94,granite33_2b,gsm8k,94,16,0.6,1,5031356,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s94/environment-5031356.json,32302.0,idle;weight_sync,0
30
+ strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s95,granite33_2b,gsm8k,95,16,0.6,1,5031357,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s95/environment-5031357.json,32302.0,idle,0
31
+ strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s96,granite33_2b,gsm8k,96,16,0.6,1,5031366,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-actor_only-s96/environment-5031366.json,32302.0,idle,0
32
+ strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s94,granite33_2b,gsm8k,94,8,0.7,1,5031393,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s94/environment-5031393.json,36302.0,idle;weight_sync,0
33
+ strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s95,granite33_2b,gsm8k,95,8,0.7,1,5031394,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s95/environment-5031394.json,36302.0,idle;weight_sync,0
34
+ strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s96,granite33_2b,gsm8k,96,8,0.7,1,5031400,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-reservation_only-s96/environment-5031400.json,36302.0,idle;weight_sync,0
35
+ strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s94,granite33_2b,gsm8k,94,16,0.7,1,5031401,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s94/environment-5031401.json,36302.0,idle,0
36
+ strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s95,granite33_2b,gsm8k,95,16,0.7,1,5031402,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s95/environment-5031402.json,36302.0,idle;weight_sync,0
37
+ strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s96,granite33_2b,gsm8k,96,16,0.7,1,5031403,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-gsm8k_long-joint-s96/environment-5031403.json,36302.0,idle;weight_sync,0
38
+ strong-factorial-granite33_2b-grpo-math_short-base-s94,granite33_2b,math,94,8,0.6,1,5031406,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-base-s94/environment-5031406.json,32710.0,idle;weight_sync,0
39
+ strong-factorial-granite33_2b-grpo-math_short-base-s95,granite33_2b,math,95,8,0.6,1,5031407,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-base-s95/environment-5031407.json,32690.0,idle;weight_sync,0
40
+ strong-factorial-granite33_2b-grpo-math_short-base-s96,granite33_2b,math,96,8,0.6,1,5031428,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-base-s96/environment-5031428.json,32690.0,idle,0
41
+ strong-factorial-granite33_2b-grpo-math_short-actor_only-s94,granite33_2b,math,94,16,0.6,1,5031429,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-actor_only-s94/environment-5031429.json,32670.0,idle,0
42
+ strong-factorial-granite33_2b-grpo-math_short-actor_only-s95,granite33_2b,math,95,16,0.6,1,5031430,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-actor_only-s95/environment-5031430.json,32670.0,idle;weight_sync,0
43
+ strong-factorial-granite33_2b-grpo-math_short-actor_only-s96,granite33_2b,math,96,16,0.6,1,5031431,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-actor_only-s96/environment-5031431.json,32670.0,idle;weight_sync,0
44
+ strong-factorial-granite33_2b-grpo-math_short-reservation_only-s94,granite33_2b,math,94,8,0.7,1,5031432,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-reservation_only-s94/environment-5031432.json,36710.0,idle,0
45
+ strong-factorial-granite33_2b-grpo-math_short-reservation_only-s95,granite33_2b,math,95,8,0.7,1,5031435,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-reservation_only-s95/environment-5031435.json,36690.0,idle;weight_sync,0
46
+ strong-factorial-granite33_2b-grpo-math_short-reservation_only-s96,granite33_2b,math,96,8,0.7,1,5031438,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-reservation_only-s96/environment-5031438.json,36690.0,idle;weight_sync,0
47
+ strong-factorial-granite33_2b-grpo-math_short-joint-s94,granite33_2b,math,94,16,0.7,1,5031439,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-joint-s94/environment-5031439.json,36670.0,idle,0
48
+ strong-factorial-granite33_2b-grpo-math_short-joint-s95,granite33_2b,math,95,16,0.7,1,5031497,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-joint-s95/environment-5031497.json,36670.0,idle,0
49
+ strong-factorial-granite33_2b-grpo-math_short-joint-s96,granite33_2b,math,96,16,0.7,1,5031498,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-granite33_2b-grpo-math_short-joint-s96/environment-5031498.json,36670.0,idle;weight_sync,0
50
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s94,phi4_mini,code_heavy_tail,94,8,0.6,1,5031694,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s94/environment-5031694.json,36020.0,rollout,0
51
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s95,phi4_mini,code_heavy_tail,95,8,0.6,1,5031695,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s95/environment-5031695.json,35958.0,rollout,0
52
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s96,phi4_mini,code_heavy_tail,96,8,0.6,1,5031700,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-base-s96/environment-5031700.json,35970.0,rollout,0
53
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s94,phi4_mini,code_heavy_tail,94,16,0.6,1,5031706,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s94/environment-5031706.json,35958.0,rollout,0
54
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s95,phi4_mini,code_heavy_tail,95,16,0.6,1,5031718,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s95/environment-5031718.json,36100.0,rollout,0
55
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s96,phi4_mini,code_heavy_tail,96,16,0.6,1,5031719,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-actor_only-s96/environment-5031719.json,36100.0,rollout,0
56
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s94,phi4_mini,code_heavy_tail,94,8,0.7,1,5031720,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s94/environment-5031720.json,40074.0,rollout,0
57
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s95,phi4_mini,code_heavy_tail,95,8,0.7,1,5031721,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s95/environment-5031721.json,39938.0,rollout,0
58
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s96,phi4_mini,code_heavy_tail,96,8,0.7,1,5031728,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-reservation_only-s96/environment-5031728.json,40062.0,rollout,0
59
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s94,phi4_mini,code_heavy_tail,94,16,0.7,1,5031729,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s94/environment-5031729.json,40052.0,rollout,0
60
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s95,phi4_mini,code_heavy_tail,95,16,0.7,1,5031730,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s95/environment-5031730.json,40142.0,rollout,0
61
+ strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s96,phi4_mini,code_heavy_tail,96,16,0.7,1,5031737,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_heavy_tail-joint-s96/environment-5031737.json,39920.0,rollout,0
62
+ strong-factorial-phi4_mini-grpo-code_standard-base-s94,phi4_mini,code_standard,94,8,0.6,1,5031745,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-base-s94/environment-5031745.json,35694.0,rollout,0
63
+ strong-factorial-phi4_mini-grpo-code_standard-base-s95,phi4_mini,code_standard,95,8,0.6,1,5031746,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-base-s95/environment-5031746.json,35958.0,rollout,0
64
+ strong-factorial-phi4_mini-grpo-code_standard-base-s96,phi4_mini,code_standard,96,8,0.6,1,5031747,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-base-s96/environment-5031747.json,35522.0,rollout,0
65
+ strong-factorial-phi4_mini-grpo-code_standard-actor_only-s94,phi4_mini,code_standard,94,16,0.6,1,5031749,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-actor_only-s94/environment-5031749.json,35720.0,rollout,0
66
+ strong-factorial-phi4_mini-grpo-code_standard-actor_only-s95,phi4_mini,code_standard,95,16,0.6,1,5031754,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-actor_only-s95/environment-5031754.json,35588.0,rollout,0
67
+ strong-factorial-phi4_mini-grpo-code_standard-actor_only-s96,phi4_mini,code_standard,96,16,0.6,1,5031755,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-actor_only-s96/environment-5031755.json,35822.0,rollout,0
68
+ strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s94,phi4_mini,code_standard,94,8,0.7,1,5031756,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s94/environment-5031756.json,39876.0,rollout,0
69
+ strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s95,phi4_mini,code_standard,95,8,0.7,1,5031759,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s95/environment-5031759.json,39708.0,rollout,0
70
+ strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s96,phi4_mini,code_standard,96,8,0.7,1,5031767,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-reservation_only-s96/environment-5031767.json,39550.0,idle;weight_sync,0
71
+ strong-factorial-phi4_mini-grpo-code_standard-joint-s94,phi4_mini,code_standard,94,16,0.7,1,5031768,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-joint-s94/environment-5031768.json,39668.0,rollout,0
72
+ strong-factorial-phi4_mini-grpo-code_standard-joint-s95,phi4_mini,code_standard,95,16,0.7,1,5031769,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-joint-s95/environment-5031769.json,39730.0,rollout,0
73
+ strong-factorial-phi4_mini-grpo-code_standard-joint-s96,phi4_mini,code_standard,96,16,0.7,1,5031772,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-code_standard-joint-s96/environment-5031772.json,39892.0,rollout,0
74
+ strong-factorial-phi4_mini-grpo-gsm8k-base-s94,phi4_mini,gsm8k,94,8,0.6,1,5031778,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-base-s94/environment-5031778.json,35518.0,idle;weight_sync,0
75
+ strong-factorial-phi4_mini-grpo-gsm8k-base-s95,phi4_mini,gsm8k,95,8,0.6,1,5031779,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-base-s95/environment-5031779.json,35518.0,idle;weight_sync,0
76
+ strong-factorial-phi4_mini-grpo-gsm8k-base-s96,phi4_mini,gsm8k,96,8,0.6,1,5031780,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-base-s96/environment-5031780.json,35518.0,idle;weight_sync,0
77
+ strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s94,phi4_mini,gsm8k,94,16,0.6,1,5031783,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s94/environment-5031783.json,35538.0,idle,0
78
+ strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s95,phi4_mini,gsm8k,95,16,0.6,1,5031791,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s95/environment-5031791.json,35518.0,idle;weight_sync,0
79
+ strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s96,phi4_mini,gsm8k,96,16,0.6,1,5031796,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-actor_only-s96/environment-5031796.json,35538.0,idle,0
80
+ strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s94,phi4_mini,gsm8k,94,8,0.7,1,5031797,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s94/environment-5031797.json,39550.0,idle;weight_sync,0
81
+ strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s95,phi4_mini,gsm8k,95,8,0.7,1,5031801,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s95/environment-5031801.json,39550.0,idle,0
82
+ strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s96,phi4_mini,gsm8k,96,8,0.7,1,5031806,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-reservation_only-s96/environment-5031806.json,39550.0,idle;weight_sync,0
83
+ strong-factorial-phi4_mini-grpo-gsm8k-joint-s94,phi4_mini,gsm8k,94,16,0.7,1,5031807,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-joint-s94/environment-5031807.json,39550.0,idle;weight_sync,0
84
+ strong-factorial-phi4_mini-grpo-gsm8k-joint-s95,phi4_mini,gsm8k,95,16,0.7,1,5031808,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-joint-s95/environment-5031808.json,39550.0,idle;weight_sync,0
85
+ strong-factorial-phi4_mini-grpo-gsm8k-joint-s96,phi4_mini,gsm8k,96,16,0.7,1,5031812,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-gsm8k-joint-s96/environment-5031812.json,39570.0,idle;weight_sync,0
86
+ strong-factorial-phi4_mini-grpo-math-base-s94,phi4_mini,math,94,8,0.6,1,5031823,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-base-s94/environment-5031823.json,36158.0,rollout,0
87
+ strong-factorial-phi4_mini-grpo-math-base-s95,phi4_mini,math,95,8,0.6,1,5031824,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-base-s95/environment-5031824.json,36086.0,rollout,0
88
+ strong-factorial-phi4_mini-grpo-math-base-s96,phi4_mini,math,96,8,0.6,1,5031826,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-base-s96/environment-5031826.json,35920.0,initializing,0
89
+ strong-factorial-phi4_mini-grpo-math-actor_only-s94,phi4_mini,math,94,16,0.6,1,5031829,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-actor_only-s94/environment-5031829.json,35938.0,rollout,0
90
+ strong-factorial-phi4_mini-grpo-math-actor_only-s95,phi4_mini,math,95,16,0.6,1,5031836,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-actor_only-s95/environment-5031836.json,36602.0,rollout,0
91
+ strong-factorial-phi4_mini-grpo-math-actor_only-s96,phi4_mini,math,96,16,0.6,1,5031837,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-actor_only-s96/environment-5031837.json,35920.0,initializing,0
92
+ strong-factorial-phi4_mini-grpo-math-reservation_only-s94,phi4_mini,math,94,8,0.7,1,5031838,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-reservation_only-s94/environment-5031838.json,40058.0,rollout,0
93
+ strong-factorial-phi4_mini-grpo-math-reservation_only-s95,phi4_mini,math,95,8,0.7,1,5031841,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-reservation_only-s95/environment-5031841.json,40016.0,initializing,0
94
+ strong-factorial-phi4_mini-grpo-math-reservation_only-s96,phi4_mini,math,96,8,0.7,1,5031856,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-reservation_only-s96/environment-5031856.json,40594.0,rollout,0
95
+ strong-factorial-phi4_mini-grpo-math-joint-s94,phi4_mini,math,94,16,0.7,1,5031857,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-joint-s94/environment-5031857.json,40160.0,rollout,0
96
+ strong-factorial-phi4_mini-grpo-math-joint-s95,phi4_mini,math,95,16,0.7,1,5031860,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-joint-s95/environment-5031860.json,40120.0,rollout,0
97
+ strong-factorial-phi4_mini-grpo-math-joint-s96,phi4_mini,math,96,16,0.7,1,5031862,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-phi4_mini-grpo-math-joint-s96/environment-5031862.json,40016.0,initializing,0
98
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s94,qwen25_3b,code_heavy_tail,94,8,0.6,1,5031981,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s94/environment-5031981.json,33430.0,rollout,0
99
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s95,qwen25_3b,code_heavy_tail,95,8,0.6,1,5031982,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s95/environment-5031982.json,33546.0,rollout,0
100
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s96,qwen25_3b,code_heavy_tail,96,8,0.6,1,5031983,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-base-s96/environment-5031983.json,35156.0,idle;weight_sync,0
101
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s94,qwen25_3b,code_heavy_tail,94,16,0.6,1,5031985,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s94/environment-5031985.json,33498.0,rollout,0
102
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s95,qwen25_3b,code_heavy_tail,95,16,0.6,1,5031987,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s95/environment-5031987.json,33524.0,rollout,0
103
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s96,qwen25_3b,code_heavy_tail,96,16,0.6,1,5031988,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-actor_only-s96/environment-5031988.json,33386.0,rollout,0
104
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s94,qwen25_3b,code_heavy_tail,94,8,0.7,1,5031989,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s94/environment-5031989.json,37418.0,rollout,0
105
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s95,qwen25_3b,code_heavy_tail,95,8,0.7,1,5031990,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s95/environment-5031990.json,37418.0,rollout,0
106
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s96,qwen25_3b,code_heavy_tail,96,8,0.7,1,5031991,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-reservation_only-s96/environment-5031991.json,37412.0,rollout,0
107
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s94,qwen25_3b,code_heavy_tail,94,16,0.7,1,5031992,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s94/environment-5031992.json,37418.0,rollout,0
108
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s95,qwen25_3b,code_heavy_tail,95,16,0.7,1,5031993,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s95/environment-5031993.json,37418.0,rollout,0
109
+ strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s96,qwen25_3b,code_heavy_tail,96,16,0.7,1,5031994,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_heavy_tail-joint-s96/environment-5031994.json,37418.0,rollout,0
110
+ strong-factorial-qwen25_3b-grpo-code_standard-base-s94,qwen25_3b,code_standard,94,8,0.6,1,5032005,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-base-s94/environment-5032005.json,33386.0,rollout,0
111
+ strong-factorial-qwen25_3b-grpo-code_standard-base-s95,qwen25_3b,code_standard,95,8,0.6,1,5032006,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-base-s95/environment-5032006.json,33298.0,rollout,0
112
+ strong-factorial-qwen25_3b-grpo-code_standard-base-s96,qwen25_3b,code_standard,96,8,0.6,1,5032007,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-base-s96/environment-5032007.json,33224.0,rollout;unknown,0
113
+ strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s94,qwen25_3b,code_standard,94,16,0.6,1,5032012,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s94/environment-5032012.json,33108.0,idle,0
114
+ strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s95,qwen25_3b,code_standard,95,16,0.6,1,5032015,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s95/environment-5032015.json,33386.0,rollout,0
115
+ strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s96,qwen25_3b,code_standard,96,16,0.6,1,5032016,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-actor_only-s96/environment-5032016.json,33108.0,idle;weight_sync,0
116
+ strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s94,qwen25_3b,code_standard,94,8,0.7,1,5032033,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s94/environment-5032033.json,37418.0,rollout,0
117
+ strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s95,qwen25_3b,code_standard,95,8,0.7,1,5032018,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s95/environment-5032018.json,37418.0,rollout,0
118
+ strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s96,qwen25_3b,code_standard,96,8,0.7,1,5032020,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-reservation_only-s96/environment-5032020.json,37418.0,rollout,0
119
+ strong-factorial-qwen25_3b-grpo-code_standard-joint-s94,qwen25_3b,code_standard,94,16,0.7,1,5032024,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-joint-s94/environment-5032024.json,37418.0,rollout,0
120
+ strong-factorial-qwen25_3b-grpo-code_standard-joint-s95,qwen25_3b,code_standard,95,16,0.7,1,5032025,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-joint-s95/environment-5032025.json,37418.0,rollout,0
121
+ strong-factorial-qwen25_3b-grpo-code_standard-joint-s96,qwen25_3b,code_standard,96,16,0.7,1,5032026,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-code_standard-joint-s96/environment-5032026.json,37346.0,rollout,0
122
+ strong-factorial-qwen25_3b-grpo-gsm8k-base-s94,qwen25_3b,gsm8k,94,8,0.6,1,5032028,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-base-s94/environment-5032028.json,33108.0,idle;weight_sync,0
123
+ strong-factorial-qwen25_3b-grpo-gsm8k-base-s95,qwen25_3b,gsm8k,95,8,0.6,1,5032029,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-base-s95/environment-5032029.json,33108.0,idle;weight_sync,0
124
+ strong-factorial-qwen25_3b-grpo-gsm8k-base-s96,qwen25_3b,gsm8k,96,8,0.6,1,5032030,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-base-s96/environment-5032030.json,33108.0,idle,0
125
+ strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s94,qwen25_3b,gsm8k,94,16,0.6,1,5032031,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s94/environment-5032031.json,33108.0,idle;weight_sync,0
126
+ strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s95,qwen25_3b,gsm8k,95,16,0.6,1,5032034,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s95/environment-5032034.json,33108.0,idle;weight_sync,0
127
+ strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s96,qwen25_3b,gsm8k,96,16,0.6,1,5032036,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-actor_only-s96/environment-5032036.json,33108.0,idle;weight_sync,0
128
+ strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s94,qwen25_3b,gsm8k,94,8,0.7,1,5032037,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s94/environment-5032037.json,37140.0,idle;weight_sync,0
129
+ strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s95,qwen25_3b,gsm8k,95,8,0.7,1,5032038,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s95/environment-5032038.json,37140.0,idle,0
130
+ strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s96,qwen25_3b,gsm8k,96,8,0.7,1,5032039,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-reservation_only-s96/environment-5032039.json,37140.0,idle,0
131
+ strong-factorial-qwen25_3b-grpo-gsm8k-joint-s94,qwen25_3b,gsm8k,94,16,0.7,1,5032040,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-joint-s94/environment-5032040.json,37140.0,idle;weight_sync,0
132
+ strong-factorial-qwen25_3b-grpo-gsm8k-joint-s95,qwen25_3b,gsm8k,95,16,0.7,1,5032041,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-joint-s95/environment-5032041.json,37140.0,idle;weight_sync,0
133
+ strong-factorial-qwen25_3b-grpo-gsm8k-joint-s96,qwen25_3b,gsm8k,96,16,0.7,1,5032042,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-gsm8k-joint-s96/environment-5032042.json,37140.0,idle;weight_sync,0
134
+ strong-factorial-qwen25_3b-grpo-math-base-s94,qwen25_3b,math,94,8,0.6,1,5032043,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-base-s94/environment-5032043.json,33446.0,idle;weight_sync,0
135
+ strong-factorial-qwen25_3b-grpo-math-base-s95,qwen25_3b,math,95,8,0.6,1,5032045,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-base-s95/environment-5032045.json,33446.0,idle,0
136
+ strong-factorial-qwen25_3b-grpo-math-base-s96,qwen25_3b,math,96,8,0.6,1,5032046,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-base-s96/environment-5032046.json,33446.0,idle;weight_sync,0
137
+ strong-factorial-qwen25_3b-grpo-math-actor_only-s94,qwen25_3b,math,94,16,0.6,1,5032047,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-actor_only-s94/environment-5032047.json,33446.0,idle;weight_sync,0
138
+ strong-factorial-qwen25_3b-grpo-math-actor_only-s95,qwen25_3b,math,95,16,0.6,1,5032051,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-actor_only-s95/environment-5032051.json,33446.0,idle,0
139
+ strong-factorial-qwen25_3b-grpo-math-actor_only-s96,qwen25_3b,math,96,16,0.6,1,5032052,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-actor_only-s96/environment-5032052.json,33446.0,idle;weight_sync,0
140
+ strong-factorial-qwen25_3b-grpo-math-reservation_only-s94,qwen25_3b,math,94,8,0.7,1,5032053,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-reservation_only-s94/environment-5032053.json,37478.0,idle;weight_sync,0
141
+ strong-factorial-qwen25_3b-grpo-math-reservation_only-s95,qwen25_3b,math,95,8,0.7,1,5032054,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-reservation_only-s95/environment-5032054.json,37478.0,idle;weight_sync,0
142
+ strong-factorial-qwen25_3b-grpo-math-reservation_only-s96,qwen25_3b,math,96,8,0.7,1,5032056,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-reservation_only-s96/environment-5032056.json,37478.0,idle;weight_sync,0
143
+ strong-factorial-qwen25_3b-grpo-math-joint-s94,qwen25_3b,math,94,16,0.7,1,5032058,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-joint-s94/environment-5032058.json,37478.0,idle,0
144
+ strong-factorial-qwen25_3b-grpo-math-joint-s95,qwen25_3b,math,95,16,0.7,1,5032059,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-joint-s95/environment-5032059.json,37478.0,idle;weight_sync,0
145
+ strong-factorial-qwen25_3b-grpo-math-joint-s96,qwen25_3b,math,96,16,0.7,1,5032060,/mnt/proj3/open-35-44/verl/rl/output/strengthening_factorial/strong-factorial-qwen25_3b-grpo-math-joint-s96/environment-5032060.json,37478.0,idle;weight_sync,0
tables/review_revision/factorial-workload-effects.csv ADDED
@@ -0,0 +1,5 @@
 
 
 
 
 
 
1
+ dataset,blocks,actor_main_effect_whole_run_mib,reservation_main_effect_whole_run_mib,interaction_whole_run_mib,actor_main_effect_actor_nvml_mib,reservation_main_effect_actor_nvml_mib,interaction_actor_nvml_mib,actor_main_effect_actor_allocated_mib,reservation_main_effect_actor_allocated_mib,interaction_actor_allocated_mib
2
+ code_heavy_tail,9,-95.44444444444444,3881.4444444444443,192.66666666666666,6267.555555555556,-343.55555555555554,-34.22222222222222,6072.907172309027,-309.62456597222223,-40.21929253472222
3
+ code_standard,9,-13.88888888888889,4064.1111111111113,-39.77777777777778,4131.888888888889,167.22222222222223,69.55555555555556,4006.3097873263887,130.11642795138889,41.73871527777778
4
+ gsm8k,9,0.0,4016.8888888888887,4.444444444444445,2072.8888888888887,38.44444444444444,117.77777777777777,2000.4676649305557,48.40424262152778,185.91959635416666
5
+ math,9,-13.11111111111111,4029.5555555555557,-74.22222222222223,3447.8888888888887,173.22222222222223,305.55555555555554,3311.3446994357637,133.652099609375,259.77088758680554
tables/review_revision/instrumentation-stage-cells.csv ADDED
@@ -0,0 +1,13 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ model_family,dataset,configuration_level,pairs,actor_update_observed_pairs,actor_update_median_difference_mib,actor_update_min_difference_mib,actor_update_max_difference_mib,actor_update_mean_duration_difference_seconds,rollout_observed_pairs,rollout_median_difference_mib,rollout_min_difference_mib,rollout_max_difference_mib,rollout_mean_duration_difference_seconds,weight_sync_observed_pairs,weight_sync_median_difference_mib,weight_sync_min_difference_mib,weight_sync_max_difference_mib,weight_sync_mean_duration_difference_seconds
2
+ granite33_2b,gsm8k,c3,3,3,-114.0,-424.0,-38.0,0.6999925033333335,3,-8.0,-100.0,6.0,-5.898574166666674,3,0.0,0.0,0.0,0.0996689066666659
3
+ granite33_2b,gsm8k,c4,3,3,-120.0,-334.0,-100.0,-0.033329876666667126,3,-96.0,-166.0,-6.0,-5.099992995666675,3,0.0,0.0,0.0,-0.5333348136666659
4
+ granite33_2b,math,c3,3,3,-60.0,-198.0,56.0,0.7666605243333313,3,-58.0,-272.0,170.0,0.9666687290000032,3,0.0,0.0,0.0,0.06667422933333118
5
+ granite33_2b,math,c4,3,3,36.0,-216.0,140.0,-2.5000091726666676,3,-18.0,-128.0,136.0,-0.7666675686666764,3,0.0,0.0,0.0,-1.3999914710000052
6
+ phi4_mini,gsm8k,c2,3,3,20.0,-1800.0,1456.0,-0.13330663366666595,3,22.0,0.0,28.0,-0.5666731779999831,3,-20.0,-502.0,502.0,1.5666608756666658
7
+ phi4_mini,gsm8k,c3,3,3,362.0,-320.0,550.0,-1.5999868283333336,3,24.0,-52.0,68.0,1.3333316553333259,3,0.0,-20.0,608.0,-0.2668878249999948
8
+ phi4_mini,math,c2,3,3,-176.0,-192.0,442.0,-0.0255855796666656,3,-46.0,-402.0,76.0,4.434521247000004,3,0.0,0.0,0.0,0.5999881566666646
9
+ phi4_mini,math,c3,3,3,398.0,280.0,798.0,0.06667501633333266,3,-330.0,-374.0,518.0,-0.6666708513333374,3,0.0,0.0,20.0,3.033333976333328
10
+ qwen25_3b,gsm8k,c3,3,3,116.0,-536.0,164.0,0.4333235646666669,3,144.0,86.0,148.0,-7.933321414999981,3,0.0,0.0,0.0,-0.7666523943333345
11
+ qwen25_3b,gsm8k,c4,3,3,400.0,0.0,404.0,0.23332970266666594,3,0.0,-50.0,38.0,-2.733356490999993,0,nan,nan,nan,nan
12
+ qwen25_3b,math,c3,3,3,60.0,-532.0,216.0,0.0333283256666661,3,-22.0,-42.0,106.0,3.299987941000005,3,0.0,0.0,0.0,1.1333318190000046
13
+ qwen25_3b,math,c4,3,0,nan,nan,nan,nan,0,nan,nan,nan,nan,0,nan,nan,nan,nan
tables/review_revision/instrumentation-stage-pairs.csv ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ pair_id,model_family,dataset,configuration_level,training_seed,first_condition,both_complete,actor_update_full_minus_external_peak_mib,actor_update_full_minus_external_sampled_seconds,rollout_full_minus_external_peak_mib,rollout_full_minus_external_sampled_seconds,weight_sync_full_minus_external_peak_mib,weight_sync_full_minus_external_sampled_seconds
2
+ qwen25_3b-gsm8k-c3-s111,qwen25_3b,gsm8k,c3,111,external,1,-536.0,0.09999533400000038,148.0,-22.00000293299998,0.0,-0.1999934680000024
3
+ qwen25_3b-gsm8k-c3-s112,qwen25_3b,gsm8k,c3,112,full,1,116.0,0.09998830100000022,144.0,0.09999270800002336,0.0,-1.3999782310000004
4
+ qwen25_3b-gsm8k-c3-s113,qwen25_3b,gsm8k,c3,113,external,1,164.0,1.099987059,86.0,-1.8999540199999885,0.0,-0.6999854840000008
5
+ qwen25_3b-gsm8k-c4-s111,qwen25_3b,gsm8k,c4,111,full,0,404.0,0.09999841900000028,0.0,-1.2000538069999962,nan,nan
6
+ qwen25_3b-gsm8k-c4-s112,qwen25_3b,gsm8k,c4,112,external,0,0.0,-1.470000010606043e-07,-50.0,-0.7000047779999861,nan,nan
7
+ qwen25_3b-gsm8k-c4-s113,qwen25_3b,gsm8k,c4,113,full,0,400.0,0.5999908359999986,38.0,-6.300010887999996,nan,nan
8
+ qwen25_3b-math-c3-s111,qwen25_3b,math,c3,111,full,1,216.0,-1.5570000000408868e-05,-22.0,5.099991562000014,0.0,2.700017894000018
9
+ qwen25_3b-math-c3-s112,qwen25_3b,math,c3,112,external,1,-532.0,0.1000068510000034,-42.0,4.799975280999991,0.0,-1.5458999998330114e-05
10
+ qwen25_3b-math-c3-s113,qwen25_3b,math,c3,113,full,1,60.0,-6.3040000046754585e-06,106.0,-3.019999990527822e-06,0.0,0.6999930219999939
11
+ qwen25_3b-math-c4-s111,qwen25_3b,math,c4,111,external,0,nan,nan,nan,nan,nan,nan
12
+ qwen25_3b-math-c4-s112,qwen25_3b,math,c4,112,full,0,nan,nan,nan,nan,nan,nan
13
+ qwen25_3b-math-c4-s113,qwen25_3b,math,c4,113,external,0,nan,nan,nan,nan,nan,nan
14
+ phi4_mini-gsm8k-c2-s111,phi4_mini,gsm8k,c2,111,full,1,-1800.0,-0.39995366599999915,0.0,-15.500012969999982,502.0,0.29995284600000005
15
+ phi4_mini-gsm8k-c2-s112,phi4_mini,gsm8k,c2,112,external,1,1456.0,-0.09996924799999984,28.0,16.29999250500002,-502.0,3.899969783999998
16
+ phi4_mini-gsm8k-c2-s113,phi4_mini,gsm8k,c2,113,full,1,20.0,0.10000301300000114,22.0,-2.499999068999987,-20.0,0.5000599969999993
17
+ phi4_mini-gsm8k-c3-s111,phi4_mini,gsm8k,c3,111,external,1,550.0,0.29999010799999937,24.0,-2.100003177000019,-20.0,-0.3995455889999935
18
+ phi4_mini-gsm8k-c3-s112,phi4_mini,gsm8k,c3,112,full,1,-320.0,-4.799954904000001,-52.0,4.899993278000004,0.0,-1.3011161869999999
19
+ phi4_mini-gsm8k-c3-s113,phi4_mini,gsm8k,c3,113,external,1,362.0,-0.29999568899999973,68.0,1.2000048649999933,608.0,0.899998301000009
20
+ phi4_mini-math-c2-s111,phi4_mini,math,c2,111,external,1,-176.0,-0.6767678170000018,-46.0,7.1000044559999935,0.0,-0.6000067390000083
21
+ phi4_mini-math-c2-s112,phi4_mini,math,c2,112,full,1,-192.0,-0.10000032899999844,-402.0,-0.09644700700000186,0.0,-5.145999994127237e-06
22
+ phi4_mini-math-c2-s113,phi4_mini,math,c2,113,external,1,442.0,0.7000114070000034,76.0,6.30000629200002,0.0,2.399976354999996
23
+ phi4_mini-math-c3-s111,phi4_mini,math,c3,111,full,1,798.0,0.10002644999999966,518.0,-3.800011090000016,0.0,10.199995591999986
24
+ phi4_mini-math-c3-s112,phi4_mini,math,c3,112,external,1,280.0,0.09999748899999794,-330.0,1.6999996629999963,20.0,0.10000169900000166
25
+ phi4_mini-math-c3-s113,phi4_mini,math,c3,113,full,1,398.0,1.1100000003594346e-06,-374.0,0.09999887300000765,0.0,-1.1999953620000046
26
+ granite33_2b-gsm8k-c3-s111,granite33_2b,gsm8k,c3,111,external,1,-38.0,-0.10001479600000085,6.0,-4.2957254209999896,0.0,-0.19998914499999998
27
+ granite33_2b-gsm8k-c3-s112,granite33_2b,gsm8k,c3,112,full,1,-114.0,0.7999894580000007,-8.0,1.0999929849999859,0.0,-0.20099514099999638
28
+ granite33_2b-gsm8k-c3-s113,granite33_2b,gsm8k,c3,113,external,1,-424.0,1.4000028480000006,-100.0,-14.49999006400002,0.0,0.6999910059999941
29
+ granite33_2b-gsm8k-c4-s111,granite33_2b,gsm8k,c4,111,full,1,-334.0,0.19999766699999988,-166.0,-13.19999414500002,0.0,-1.5999894099999974
30
+ granite33_2b-gsm8k-c4-s112,granite33_2b,gsm8k,c4,112,external,1,-100.0,-0.29999027400000067,-96.0,-0.7999950870000276,0.0,-1.0000123240000018
31
+ granite33_2b-gsm8k-c4-s113,granite33_2b,gsm8k,c4,113,full,1,-120.0,2.9769999994044838e-06,-6.0,-1.2999897549999773,0.0,0.9999972930000016
32
+ granite33_2b-math-c3-s111,granite33_2b,math,c3,111,full,1,56.0,0.40000697999999524,-58.0,-0.09998372500000485,0.0,-1.433000004880114e-06
33
+ granite33_2b-math-c3-s112,granite33_2b,math,c3,112,external,1,-60.0,0.49999715000000045,170.0,0.1999911440000126,0.0,0.20000511599999982
34
+ granite33_2b-math-c3-s113,granite33_2b,math,c3,113,full,1,-198.0,1.3999774429999983,-272.0,2.799998768000002,0.0,1.900499999862859e-05
35
+ granite33_2b-math-c4-s111,granite33_2b,math,c4,111,external,1,-216.0,0.19998975800000363,-18.0,-0.2999914950000111,0.0,-4.399985957000007
36
+ granite33_2b-math-c4-s112,granite33_2b,math,c4,112,full,1,140.0,-7.700003003000005,136.0,0.4999893289999946,0.0,-1.0999998180000023
37
+ granite33_2b-math-c4-s113,granite33_2b,math,c4,113,external,1,36.0,-1.4273000000564195e-05,-128.0,-2.5000005400000127,0.0,1.300011361999994
tables/review_revision/instrumentation-stage-trials.csv ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ experiment_id,pair_id,model_family,dataset,configuration_level,training_seed,condition,period,success,terminal_phase,actor_update_observed_peak_mib,actor_update_finished,actor_update_peak_mib,actor_update_sampled_seconds,rollout_observed_peak_mib,rollout_finished,rollout_peak_mib,rollout_sampled_seconds,weight_sync_observed_peak_mib,weight_sync_finished,weight_sync_peak_mib,weight_sync_sampled_seconds
2
+ rev-instrumentation-qwen25_3b-gsm8k-c3-s111-external,qwen25_3b-gsm8k-c3-s111,qwen25_3b,gsm8k,c3,111,external,1,1,idle,17264.0,1,17264.0,3.6000030769999998,36692.0,1,36692.0,44.19999773099998,37140.0,1,37140.0,12.599994239
3
+ rev-instrumentation-qwen25_3b-gsm8k-c3-s111-full,qwen25_3b-gsm8k-c3-s111,qwen25_3b,gsm8k,c3,111,full,2,1,idle,16728.0,1,16728.0,3.699998411,36840.0,1,36840.0,22.199994798000002,37140.0,1,37140.0,12.400000770999998
4
+ rev-instrumentation-qwen25_3b-gsm8k-c3-s112-full,qwen25_3b-gsm8k-c3-s112,qwen25_3b,gsm8k,c3,112,full,1,1,idle,16804.0,1,16804.0,3.699994821,36836.0,1,36836.0,21.99999865700001,37140.0,1,37140.0,10.900009005000003
5
+ rev-instrumentation-qwen25_3b-gsm8k-c3-s112-external,qwen25_3b-gsm8k-c3-s112,qwen25_3b,gsm8k,c3,112,external,2,1,idle,16688.0,1,16688.0,3.60000652,36692.0,1,36692.0,21.900005948999986,37140.0,1,37140.0,12.299987236000003
6
+ rev-instrumentation-qwen25_3b-gsm8k-c3-s113-external,qwen25_3b-gsm8k-c3-s113,qwen25_3b,gsm8k,c3,113,external,1,1,idle,16464.0,1,16464.0,3.600008538999999,36692.0,1,36692.0,27.699993011999975,37140.0,1,37140.0,13.799992862999998
7
+ rev-instrumentation-qwen25_3b-gsm8k-c3-s113-full,qwen25_3b-gsm8k-c3-s113,qwen25_3b,gsm8k,c3,113,full,2,1,idle,16628.0,1,16628.0,4.699995597999999,36778.0,1,36778.0,25.800038991999987,37140.0,1,37140.0,13.100007378999997
8
+ rev-instrumentation-qwen25_3b-gsm8k-c4-s111-full,qwen25_3b-gsm8k-c4-s111,qwen25_3b,gsm8k,c4,111,full,1,0,weight_sync,16948.0,1,16948.0,3.7000109620000003,40796.0,1,40796.0,21.999991812000008,40536.0,0,nan,nan
9
+ rev-instrumentation-qwen25_3b-gsm8k-c4-s111-external,qwen25_3b-gsm8k-c4-s111,qwen25_3b,gsm8k,c4,111,external,2,0,weight_sync,16544.0,1,16544.0,3.600012543,40796.0,1,40796.0,23.200045619000004,40536.0,0,nan,nan
10
+ rev-instrumentation-qwen25_3b-gsm8k-c4-s112-external,qwen25_3b-gsm8k-c4-s112,qwen25_3b,gsm8k,c4,112,external,1,0,weight_sync,16808.0,1,16808.0,3.5999970670000008,40932.0,1,40932.0,24.79999961899999,40536.0,0,nan,nan
11
+ rev-instrumentation-qwen25_3b-gsm8k-c4-s112-full,qwen25_3b-gsm8k-c4-s112,qwen25_3b,gsm8k,c4,112,full,2,0,weight_sync,16808.0,1,16808.0,3.5999969199999997,40882.0,1,40882.0,24.099994841000004,40536.0,0,nan,nan
12
+ rev-instrumentation-qwen25_3b-gsm8k-c4-s113-full,qwen25_3b-gsm8k-c4-s113,qwen25_3b,gsm8k,c4,113,full,1,0,weight_sync,16948.0,1,16948.0,4.299999068,40834.0,1,40834.0,25.699995540999986,40536.0,0,nan,nan
13
+ rev-instrumentation-qwen25_3b-gsm8k-c4-s113-external,qwen25_3b-gsm8k-c4-s113,qwen25_3b,gsm8k,c4,113,external,2,0,weight_sync,16548.0,1,16548.0,3.7000082320000014,40796.0,1,40796.0,32.00000642899998,40536.0,0,nan,nan
14
+ rev-instrumentation-qwen25_3b-math-c3-s111-full,qwen25_3b-math-c3-s111,qwen25_3b,math,c3,111,full,1,1,idle,19262.0,1,19262.0,6.999995215000002,37256.0,1,37256.0,30.499989781999993,37478.0,1,37478.0,16.40000418800001
15
+ rev-instrumentation-qwen25_3b-math-c3-s111-external,qwen25_3b-math-c3-s111,qwen25_3b,math,c3,111,external,2,1,idle,19046.0,1,19046.0,7.000010785000002,37278.0,1,37278.0,25.39999821999998,37478.0,1,37478.0,13.699986293999993
16
+ rev-instrumentation-qwen25_3b-math-c3-s112-external,qwen25_3b-math-c3-s112,qwen25_3b,math,c3,112,external,1,1,idle,19338.0,1,19338.0,7.199996558999999,37266.0,1,37266.0,21.99999269100002,37478.0,1,37478.0,11.000014377999998
17
+ rev-instrumentation-qwen25_3b-math-c3-s112-full,qwen25_3b-math-c3-s112,qwen25_3b,math,c3,112,full,2,1,idle,18806.0,1,18806.0,7.300003410000002,37224.0,1,37224.0,26.79996797200001,37478.0,1,37478.0,10.999998919
18
+ rev-instrumentation-qwen25_3b-math-c3-s113-full,qwen25_3b-math-c3-s113,qwen25_3b,math,c3,113,full,1,1,idle,19326.0,1,19326.0,6.8999944019999955,37290.0,1,37290.0,21.999996129000014,37478.0,1,37478.0,11.399992062999997
19
+ rev-instrumentation-qwen25_3b-math-c3-s113-external,qwen25_3b-math-c3-s113,qwen25_3b,math,c3,113,external,2,1,idle,19266.0,1,19266.0,6.900000706,37184.0,1,37184.0,21.999999149000004,37478.0,1,37478.0,10.699999041000003
20
+ rev-instrumentation-qwen25_3b-math-c4-s111-external,qwen25_3b-math-c4-s111,qwen25_3b,math,c4,111,external,1,0,initializing,nan,0,nan,nan,nan,0,nan,nan,nan,0,nan,nan
21
+ rev-instrumentation-qwen25_3b-math-c4-s111-full,qwen25_3b-math-c4-s111,qwen25_3b,math,c4,111,full,2,0,initializing,nan,0,nan,nan,nan,0,nan,nan,nan,0,nan,nan
22
+ rev-instrumentation-qwen25_3b-math-c4-s112-full,qwen25_3b-math-c4-s112,qwen25_3b,math,c4,112,full,1,0,initializing,nan,0,nan,nan,nan,0,nan,nan,nan,0,nan,nan
23
+ rev-instrumentation-qwen25_3b-math-c4-s112-external,qwen25_3b-math-c4-s112,qwen25_3b,math,c4,112,external,2,0,initializing,nan,0,nan,nan,nan,0,nan,nan,nan,0,nan,nan
24
+ rev-instrumentation-qwen25_3b-math-c4-s113-external,qwen25_3b-math-c4-s113,qwen25_3b,math,c4,113,external,1,0,initializing,nan,0,nan,nan,nan,0,nan,nan,nan,0,nan,nan
25
+ rev-instrumentation-qwen25_3b-math-c4-s113-full,qwen25_3b-math-c4-s113,qwen25_3b,math,c4,113,full,2,0,initializing,nan,0,nan,nan,nan,0,nan,nan,nan,0,nan,nan
26
+ rev-instrumentation-phi4_mini-gsm8k-c2-s111-full,phi4_mini-gsm8k-c2-s111,phi4_mini,gsm8k,c2,111,full,1,1,idle,16290.0,1,16290.0,4.000007732,35316.0,1,35316.0,24.199989272000014,35518.0,1,35518.0,11.399988069
27
+ rev-instrumentation-phi4_mini-gsm8k-c2-s111-external,phi4_mini-gsm8k-c2-s111,phi4_mini,gsm8k,c2,111,external,2,1,idle,18090.0,1,18090.0,4.399961397999999,35316.0,1,35316.0,39.700002242,35016.0,1,35016.0,11.100035223
28
+ rev-instrumentation-phi4_mini-gsm8k-c2-s112-external,phi4_mini-gsm8k-c2-s112,phi4_mini,gsm8k,c2,112,external,1,1,idle,16786.0,1,16786.0,4.199974306,35314.0,1,35314.0,25.099997973000008,35518.0,1,35518.0,10.500023921000004
29
+ rev-instrumentation-phi4_mini-gsm8k-c2-s112-full,phi4_mini-gsm8k-c2-s112,phi4_mini,gsm8k,c2,112,full,2,1,idle,18242.0,1,18242.0,4.100005058,35342.0,1,35342.0,41.39999047800003,35016.0,1,35016.0,14.399993705000002
30
+ rev-instrumentation-phi4_mini-gsm8k-c2-s113-full,phi4_mini-gsm8k-c2-s113,phi4_mini,gsm8k,c2,113,full,1,1,idle,16702.0,1,16702.0,4.100004082000001,35336.0,1,35336.0,21.299999607000004,35518.0,1,35518.0,11.600057552999997
31
+ rev-instrumentation-phi4_mini-gsm8k-c2-s113-external,phi4_mini-gsm8k-c2-s113,phi4_mini,gsm8k,c2,113,external,2,1,idle,16682.0,1,16682.0,4.000001069,35314.0,1,35314.0,23.79999867599999,35538.0,1,35538.0,11.099997555999998
32
+ rev-instrumentation-phi4_mini-gsm8k-c3-s111-external,phi4_mini-gsm8k-c3-s111,phi4_mini,gsm8k,c3,111,external,1,1,idle,18898.0,1,18898.0,2.60001284,39346.0,1,39346.0,26.399996564000016,39570.0,1,39570.0,12.499993243999997
33
+ rev-instrumentation-phi4_mini-gsm8k-c3-s111-full,phi4_mini-gsm8k-c3-s111,phi4_mini,gsm8k,c3,111,full,2,1,idle,19448.0,1,19448.0,2.9000029479999996,39370.0,1,39370.0,24.299993386999997,39550.0,1,39550.0,12.100447655000004
34
+ rev-instrumentation-phi4_mini-gsm8k-c3-s112-full,phi4_mini-gsm8k-c3-s112,phi4_mini,gsm8k,c3,112,full,1,1,idle,18938.0,1,18938.0,3.000032634000001,39344.0,1,39344.0,26.999992674000005,39550.0,1,39550.0,12.199994922
35
+ rev-instrumentation-phi4_mini-gsm8k-c3-s112-external,phi4_mini-gsm8k-c3-s112,phi4_mini,gsm8k,c3,112,external,2,1,idle,19258.0,1,19258.0,7.799987538000002,39396.0,1,39396.0,22.099999396,39550.0,1,39550.0,13.501111109
36
+ rev-instrumentation-phi4_mini-gsm8k-c3-s113-external,phi4_mini-gsm8k-c3-s113,phi4_mini,gsm8k,c3,113,external,1,1,idle,19174.0,1,19174.0,2.9999977039999997,39348.0,1,39348.0,24.29999118600001,38942.0,1,38942.0,14.300000674999998
37
+ rev-instrumentation-phi4_mini-gsm8k-c3-s113-full,phi4_mini-gsm8k-c3-s113,phi4_mini,gsm8k,c3,113,full,2,1,idle,19536.0,1,19536.0,2.700002015,39416.0,1,39416.0,25.499996051000004,39550.0,1,39550.0,15.199998976000007
38
+ rev-instrumentation-phi4_mini-math-c2-s111-external,phi4_mini-math-c2-s111,phi4_mini,math,c2,111,external,1,1,idle,19034.0,1,19034.0,8.099997288000003,36102.0,1,36102.0,23.89999035300001,35746.0,1,35746.0,17.500011361000002
39
+ rev-instrumentation-phi4_mini-math-c2-s111-full,phi4_mini-math-c2-s111,phi4_mini,math,c2,111,full,2,1,idle,18858.0,1,18858.0,7.423229471000001,36056.0,1,36056.0,30.999994809000004,35746.0,1,35746.0,16.900004621999994
40
+ rev-instrumentation-phi4_mini-math-c2-s112-full,phi4_mini-math-c2-s112,phi4_mini,math,c2,112,full,1,1,idle,18884.0,1,18884.0,7.999994032000004,36096.0,1,36096.0,23.40353814,35746.0,1,35746.0,10.100000984000005
41
+ rev-instrumentation-phi4_mini-math-c2-s112-external,phi4_mini-math-c2-s112,phi4_mini,math,c2,112,external,2,1,idle,19076.0,1,19076.0,8.099994361000002,36498.0,1,36498.0,23.499985147,35746.0,1,35746.0,10.100006129999999
42
+ rev-instrumentation-phi4_mini-math-c2-s113-external,phi4_mini-math-c2-s113,phi4_mini,math,c2,113,external,1,1,idle,18676.0,1,18676.0,8.599998342,35794.0,1,35794.0,28.899988757000003,35746.0,1,35746.0,11.200008775000002
43
+ rev-instrumentation-phi4_mini-math-c2-s113-full,phi4_mini-math-c2-s113,phi4_mini,math,c2,113,full,2,1,idle,19118.0,1,19118.0,9.300009749000003,35870.0,1,35870.0,35.19999504900002,35746.0,1,35746.0,13.599985129999999
44
+ rev-instrumentation-phi4_mini-math-c3-s111-full,phi4_mini-math-c3-s111,phi4_mini,math,c3,111,full,1,1,idle,24056.0,1,24056.0,6.500001836999999,40386.0,1,40386.0,25.10000059699999,39842.0,1,39842.0,24.999988267999985
45
+ rev-instrumentation-phi4_mini-math-c3-s111-external,phi4_mini-math-c3-s111,phi4_mini,math,c3,111,external,2,1,idle,23258.0,1,23258.0,6.399975387,39868.0,1,39868.0,28.900011687000006,39842.0,1,39842.0,14.799992675999999
46
+ rev-instrumentation-phi4_mini-math-c3-s112-external,phi4_mini-math-c3-s112,phi4_mini,math,c3,112,external,1,1,idle,22884.0,1,22884.0,6.199998520000001,40218.0,1,40218.0,26.10000059400001,39842.0,1,39842.0,14.600000881000003
47
+ rev-instrumentation-phi4_mini-math-c3-s112-full,phi4_mini-math-c3-s112,phi4_mini,math,c3,112,full,2,1,idle,23164.0,1,23164.0,6.299996008999999,39888.0,1,39888.0,27.800000257000008,39862.0,1,39862.0,14.700002580000005
48
+ rev-instrumentation-phi4_mini-math-c3-s113-full,phi4_mini-math-c3-s113,phi4_mini,math,c3,113,full,1,1,idle,23976.0,1,23976.0,6.200006213000001,39944.0,1,39944.0,23.599997183000003,39842.0,1,39842.0,10.699994903999997
49
+ rev-instrumentation-phi4_mini-math-c3-s113-external,phi4_mini-math-c3-s113,phi4_mini,math,c3,113,external,2,1,idle,23578.0,1,23578.0,6.2000051030000005,40318.0,1,40318.0,23.499998309999995,39842.0,1,39842.0,11.899990266000001
50
+ rev-instrumentation-granite33_2b-gsm8k-c3-s111-external,granite33_2b-gsm8k-c3-s111,granite33_2b,gsm8k,c3,111,external,1,1,idle,12942.0,1,12942.0,4.200009321000001,35922.0,1,35922.0,23.700054005,36302.0,1,36302.0,11.899998323999998
51
+ rev-instrumentation-granite33_2b-gsm8k-c3-s111-full,granite33_2b-gsm8k-c3-s111,granite33_2b,gsm8k,c3,111,full,2,1,idle,12904.0,1,12904.0,4.0999945250000005,35928.0,1,35928.0,19.40432858400001,36302.0,1,36302.0,11.700009178999998
52
+ rev-instrumentation-granite33_2b-gsm8k-c3-s112-full,granite33_2b-gsm8k-c3-s112,granite33_2b,gsm8k,c3,112,full,1,1,idle,12738.0,1,12738.0,4.899997518000001,35970.0,1,35970.0,24.699994223999973,36302.0,1,36302.0,11.900008707000001
53
+ rev-instrumentation-granite33_2b-gsm8k-c3-s112-external,granite33_2b-gsm8k-c3-s112,granite33_2b,gsm8k,c3,112,external,2,1,idle,12852.0,1,12852.0,4.10000806,35978.0,1,35978.0,23.600001238999987,36302.0,1,36302.0,12.101003847999998
54
+ rev-instrumentation-granite33_2b-gsm8k-c3-s113-external,granite33_2b-gsm8k-c3-s113,granite33_2b,gsm8k,c3,113,external,1,1,idle,13402.0,1,13402.0,4.0999968959999995,35984.0,1,35984.0,36.199995360000024,36302.0,1,36302.0,11.300009012000002
55
+ rev-instrumentation-granite33_2b-gsm8k-c3-s113-full,granite33_2b-gsm8k-c3-s113,granite33_2b,gsm8k,c3,113,full,2,1,idle,12978.0,1,12978.0,5.499999744,35884.0,1,35884.0,21.700005296000004,36302.0,1,36302.0,12.000000017999996
56
+ rev-instrumentation-granite33_2b-gsm8k-c4-s111-full,granite33_2b-gsm8k-c4-s111,granite33_2b,gsm8k,c4,111,full,1,1,idle,13124.0,1,13124.0,4.899990721,39956.0,1,39956.0,23.099996708,40382.0,1,40382.0,11.800011794
57
+ rev-instrumentation-granite33_2b-gsm8k-c4-s111-external,granite33_2b-gsm8k-c4-s111,granite33_2b,gsm8k,c4,111,external,2,1,idle,13458.0,1,13458.0,4.699993054,40122.0,1,40122.0,36.29999085300002,40382.0,1,40382.0,13.400001203999997
58
+ rev-instrumentation-granite33_2b-gsm8k-c4-s112-external,granite33_2b-gsm8k-c4-s112,granite33_2b,gsm8k,c4,112,external,1,1,idle,13618.0,1,13618.0,4.299994839,40058.0,1,40058.0,35.99999251200002,40382.0,1,40382.0,11.300008894
59
+ rev-instrumentation-granite33_2b-gsm8k-c4-s112-full,granite33_2b-gsm8k-c4-s112,granite33_2b,gsm8k,c4,112,full,2,1,idle,13518.0,1,13518.0,4.000004564999999,39962.0,1,39962.0,35.19999742499999,40382.0,1,40382.0,10.299996569999998
60
+ rev-instrumentation-granite33_2b-gsm8k-c4-s113-full,granite33_2b-gsm8k-c4-s113,granite33_2b,gsm8k,c4,113,full,1,1,idle,13018.0,1,13018.0,4.10000258,40004.0,1,40004.0,24.099998876000015,40382.0,1,40382.0,13.099995767000001
61
+ rev-instrumentation-granite33_2b-gsm8k-c4-s113-external,granite33_2b-gsm8k-c4-s113,granite33_2b,gsm8k,c4,113,external,2,1,idle,13138.0,1,13138.0,4.0999996030000005,40010.0,1,40010.0,25.399988630999992,40382.0,1,40382.0,12.099998474
62
+ rev-instrumentation-granite33_2b-math-c3-s111-full,granite33_2b-math-c3-s111,granite33_2b,math,c3,111,full,1,1,idle,14488.0,1,14488.0,8.800000068999998,36314.0,1,36314.0,21.20001156899999,36670.0,1,36670.0,10.599999552999998
63
+ rev-instrumentation-granite33_2b-math-c3-s111-external,granite33_2b-math-c3-s111,granite33_2b,math,c3,111,external,2,1,idle,14432.0,1,14432.0,8.399993089000002,36372.0,1,36372.0,21.299995293999995,36670.0,1,36670.0,10.600000986000003
64
+ rev-instrumentation-granite33_2b-math-c3-s112-external,granite33_2b-math-c3-s112,granite33_2b,math,c3,112,external,1,1,idle,14550.0,1,14550.0,8.19999913,36372.0,1,36372.0,22.300004428000005,36670.0,1,36670.0,12.400000338999998
65
+ rev-instrumentation-granite33_2b-math-c3-s112-full,granite33_2b-math-c3-s112,granite33_2b,math,c3,112,full,2,1,idle,14490.0,1,14490.0,8.69999628,36542.0,1,36542.0,22.499995572000017,36670.0,1,36670.0,12.600005454999998
66
+ rev-instrumentation-granite33_2b-math-c3-s113-full,granite33_2b-math-c3-s113,granite33_2b,math,c3,113,full,1,1,idle,14076.0,1,14076.0,8.899989332999999,36326.0,1,36326.0,23.999994846000003,36670.0,1,36670.0,12.400007553
67
+ rev-instrumentation-granite33_2b-math-c3-s113-external,granite33_2b-math-c3-s113,granite33_2b,math,c3,113,external,2,1,idle,14274.0,1,14274.0,7.500011890000001,36598.0,1,36598.0,21.199996078,36670.0,1,36670.0,12.399988548000001
68
+ rev-instrumentation-granite33_2b-math-c4-s111-external,granite33_2b-math-c4-s111,granite33_2b,math,c4,111,external,1,1,idle,14588.0,1,14588.0,7.700006857999998,40608.0,1,40608.0,22.39998735600002,40750.0,1,40750.0,14.699986780000007
69
+ rev-instrumentation-granite33_2b-math-c4-s111-full,granite33_2b-math-c4-s111,granite33_2b,math,c4,111,full,2,1,idle,14372.0,1,14372.0,7.899996616000002,40590.0,1,40590.0,22.099995861000007,40750.0,1,40750.0,10.300000823
70
+ rev-instrumentation-granite33_2b-math-c4-s112-full,granite33_2b-math-c4-s112,granite33_2b,math,c4,112,full,1,1,idle,14432.0,1,14432.0,7.599994585,40544.0,1,40544.0,24.299990707000003,40750.0,1,40750.0,11.600004309999997
71
+ rev-instrumentation-granite33_2b-math-c4-s112-external,granite33_2b-math-c4-s112,granite33_2b,math,c4,112,external,2,1,idle,14292.0,1,14292.0,15.299997588000005,40408.0,1,40408.0,23.800001378000008,40750.0,1,40750.0,12.700004128
72
+ rev-instrumentation-granite33_2b-math-c4-s113-external,granite33_2b-math-c4-s113,granite33_2b,math,c4,113,external,1,1,idle,14412.0,1,14412.0,7.500008669000001,40546.0,1,40546.0,23.799996231,40750.0,1,40750.0,9.299989296000007
73
+ rev-instrumentation-granite33_2b-math-c4-s113-full,granite33_2b-math-c4-s113,granite33_2b,math,c4,113,full,2,1,idle,14448.0,1,14448.0,7.499994396000001,40418.0,1,40418.0,21.29999569099999,40750.0,1,40750.0,10.600000658
tables/review_revision/long-run-exposure.csv ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ study,model_family,dataset,completed_processes,batches_with_length_metrics,cap_reaching_batches,positive_gradient_steps,steps_with_gradient_metric,train_examples_after_filter_min,train_examples_after_filter_max,fixed_order_distinct_train_examples_derived_min,fixed_order_distinct_train_examples_derived_max,fixed_order_wraps_derived_min,fixed_order_wraps_derived_max,prompt_mean_equal_process,prompt_maximum_observed,response_mean_equal_process,response_maximum_observed
2
+ followup_100_step,granite33_2b,gsm8k,3,300,107,300,300,7473,7473,3200,3200,0,0,141.3315625,296.0,264.87463541666665,2048.0
3
+ followup_100_step,granite33_2b,math,3,300,300,300,300,7395,7395,6400,6400,0,0,144.96125,512.0,404.74765625,512.0
4
+ followup_100_step,phi4_mini,gsm8k,3,300,45,300,300,7473,7473,3200,3200,0,0,73.285,200.0,215.92270833333333,2048.0
5
+ followup_100_step,phi4_mini,math,3,300,300,300,300,7434,7434,6400,6400,0,0,83.59796875,498.0,408.2981770833333,512.0
6
+ original_100_step,qwen25_1p5b,math,6,600,600,600,600,7500,7500,7424,7424,1,1,119.669453125,1698.0,536.3655361938477,1024.0
7
+ original_100_step,qwen25_3b,gsm8k,6,600,50,600,600,7473,7473,7424,7424,1,1,103.5921875,256.0,266.59979443868,1024.0
tables/review_revision/original-temporal-case-effects.csv ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ model_family,dataset,completed_processes,maximum_later_step_increase_mib_mean,maximum_later_step_increase_mib_min,maximum_later_step_increase_mib_max,explicit_phase_excess_mib_mean,explicit_phase_excess_mib_min,explicit_phase_excess_mib_max,phase_peak_validation_mib_mean,phase_peak_validation_mib_min,phase_peak_validation_mib_max,phase_peak_checkpoint_mib_mean,phase_peak_checkpoint_mib_min,phase_peak_checkpoint_mib_max
2
+ qwen25_1p5b,math,6,2083.0,1480.0,2448.0,-2740.6666666666665,-3440.0,-2070.0,29463.0,26420.0,32514.0,13893.333333333334,13656.0,14396.0
3
+ qwen25_3b,gsm8k,6,1767.3333333333333,1464.0,2168.0,-1963.0,-2508.0,-1722.0,33027.666666666664,29974.0,36162.0,19522.666666666668,19406.0,19726.0
tables/review_revision/realized-workload-cases.csv ADDED
@@ -0,0 +1,53 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ study,model_family,dataset,retained_processes,completed_processes,processes_with_batch_length_metrics,failed_processes_with_batch_length_metrics,recorded_batches,cap_reaching_batches,positive_gradient_steps,steps_with_gradient_metric,original_train_examples_min,original_train_examples_max,selected_before_filter_min,selected_before_filter_max,train_examples_after_filter_min,train_examples_after_filter_max,filtered_train_examples_min,filtered_train_examples_max,fixed_order_distinct_train_examples_derived_min,fixed_order_distinct_train_examples_derived_max,fixed_order_wraps_derived_min,fixed_order_wraps_derived_max,prompt_mean_equal_process,prompt_maximum_observed,response_mean_equal_process,response_maximum_observed
2
+ boundary,granite33_2b,code_heavy_tail,18,15,15,0,15,1,15,15,64,64,64,64,63,63,1,1,0,32,0,0,806.6851684570313,902.0,296.03160400390624,1024.0
3
+ boundary,granite33_2b,code_standard,18,15,15,0,15,1,15,15,64,64,64,64,47,47,17,17,0,32,0,0,491.78125,588.0,227.671875,1024.0
4
+ boundary,granite33_2b,gsm8k,18,15,15,0,15,5,15,15,7473,7473,128,128,128,128,0,0,0,32,0,0,146.28125,198.0,284.76770833333336,2048.0
5
+ boundary,granite33_2b,math,18,15,15,0,15,15,15,15,7500,7500,256,256,252,252,4,4,0,64,0,0,132.296875,460.0,376.3109375,512.0
6
+ boundary,phi4_mini,code_heavy_tail,18,12,12,0,12,12,12,12,64,64,32,32,32,32,0,0,0,32,0,0,664.15625,746.0,412.0481770833333,1024.0
7
+ boundary,phi4_mini,code_standard,18,12,12,0,12,12,12,12,64,64,32,32,32,32,0,0,0,32,0,0,378.5,480.0,326.3697916666667,1024.0
8
+ boundary,phi4_mini,gsm8k,18,12,12,0,12,2,12,12,7473,7473,32,32,32,32,0,0,0,32,0,0,78.09375,121.0,225.77473958333334,2048.0
9
+ boundary,phi4_mini,math,18,12,12,0,12,12,12,12,7500,7500,64,64,64,64,0,0,0,64,0,0,76.03125,401.0,367.1842447916667,512.0
10
+ boundary,qwen25_3b,code_heavy_tail,18,12,12,0,12,7,12,12,64,64,32,32,32,32,0,0,0,32,0,0,720.1875,772.0,351.59765625,1024.0
11
+ boundary,qwen25_3b,code_standard,18,12,12,0,12,0,12,12,64,64,32,32,32,32,0,0,0,32,0,0,420.84375,518.0,285.765625,958.0
12
+ boundary,qwen25_3b,gsm8k,18,12,12,0,12,0,12,12,7473,7473,32,32,32,32,0,0,0,32,0,0,108.0,154.0,298.3411458333333,830.0
13
+ boundary,qwen25_3b,math,18,12,12,0,12,12,12,12,7500,7500,64,64,64,64,0,0,0,64,0,0,104.203125,437.0,391.986328125,512.0
14
+ factorial,granite33_2b,code_heavy_tail,12,12,12,0,12,2,12,12,64,64,64,64,63,63,1,1,32,32,0,0,806.4375,902.0,293.1184895833333,1024.0
15
+ factorial,granite33_2b,code_standard,12,12,12,0,12,0,12,12,64,64,64,64,47,47,17,17,32,32,0,0,491.78125,588.0,227.52604166666666,812.0
16
+ factorial,granite33_2b,gsm8k,12,12,12,0,12,5,12,12,7473,7473,128,128,128,128,0,0,32,32,0,0,146.28125,198.0,286.1901041666667,2048.0
17
+ factorial,granite33_2b,math,12,12,12,0,12,12,12,12,7500,7500,256,256,252,252,4,4,64,64,0,0,132.296875,460.0,375.158203125,512.0
18
+ factorial,phi4_mini,code_heavy_tail,12,12,12,0,12,12,12,12,64,64,32,32,32,32,0,0,32,32,0,0,664.15625,746.0,416.234375,1024.0
19
+ factorial,phi4_mini,code_standard,12,12,12,0,12,12,12,12,64,64,32,32,32,32,0,0,32,32,0,0,378.5,480.0,327.8723958333333,1024.0
20
+ factorial,phi4_mini,gsm8k,12,12,12,0,12,2,12,12,7473,7473,32,32,32,32,0,0,32,32,0,0,78.09375,121.0,229.33463541666666,2048.0
21
+ factorial,phi4_mini,math,12,12,12,0,12,12,12,12,7500,7500,64,64,64,64,0,0,64,64,0,0,76.03125,401.0,367.9251302083333,512.0
22
+ factorial,qwen25_3b,code_heavy_tail,12,12,12,0,12,8,12,12,64,64,32,32,32,32,0,0,32,32,0,0,720.1875,772.0,356.1979166666667,1024.0
23
+ factorial,qwen25_3b,code_standard,12,12,12,0,12,4,12,12,64,64,32,32,32,32,0,0,32,32,0,0,420.84375,518.0,295.5325520833333,1024.0
24
+ factorial,qwen25_3b,gsm8k,12,12,12,0,12,0,12,12,7473,7473,32,32,32,32,0,0,32,32,0,0,108.0,154.0,298.5182291666667,758.0
25
+ factorial,qwen25_3b,math,12,12,12,0,12,12,12,12,7500,7500,64,64,64,64,0,0,64,64,0,0,104.203125,437.0,391.283203125,512.0
26
+ followup_100_step,granite33_2b,gsm8k,6,6,6,0,303,107,303,303,7473,7473,7473,7473,7473,7473,0,0,32,3200,0,0,143.80640625,296.0,265.783671875,2048.0
27
+ followup_100_step,granite33_2b,math,6,6,6,0,303,303,303,303,7500,7500,7500,7500,7395,7395,105,105,64,6400,0,0,138.6290625,512.0,392.526171875,512.0
28
+ followup_100_step,phi4_mini,gsm8k,6,6,6,0,303,45,303,303,7473,7473,7473,7473,7473,7473,0,0,32,3200,0,0,75.689375,200.0,218.9978125,2048.0
29
+ followup_100_step,phi4_mini,math,6,6,6,0,303,303,303,303,7500,7500,7500,7500,7434,7434,66,66,64,6400,0,0,79.814609375,498.0,387.6087239583333,512.0
30
+ gpu_count,granite33_2b,gsm8k,9,9,9,0,9,2,9,9,7473,7473,128,128,128,128,0,0,32,32,0,0,146.28125,198.0,276.6736111111111,2048.0
31
+ gpu_count,granite33_2b,math,9,9,9,0,9,9,9,9,7500,7500,256,256,252,252,4,4,64,64,0,0,132.296875,460.0,378.29253472222223,512.0
32
+ gpu_count,phi4_mini,gsm8k,9,9,9,0,9,1,9,9,7473,7473,32,32,32,32,0,0,32,32,0,0,78.09375,121.0,223.75868055555554,2048.0
33
+ gpu_count,phi4_mini,math,9,9,9,0,9,9,9,9,7500,7500,64,64,64,64,0,0,64,64,0,0,76.03125,401.0,367.32899305555554,512.0
34
+ gpu_count,qwen25_3b,gsm8k,9,9,9,0,9,0,9,9,7473,7473,32,32,32,32,0,0,32,32,0,0,108.0,154.0,298.7083333333333,758.0
35
+ gpu_count,qwen25_3b,math,9,9,9,0,9,9,9,9,7500,7500,64,64,64,64,0,0,64,64,0,0,104.203125,437.0,390.7482638888889,512.0
36
+ instrumentation,granite33_2b,gsm8k,12,12,12,0,12,4,12,12,7473,7473,128,128,128,128,0,0,32,32,0,0,146.28125,198.0,283.26171875,2048.0
37
+ instrumentation,granite33_2b,math,12,12,12,0,12,12,12,12,7500,7500,256,256,252,252,4,4,64,64,0,0,132.296875,460.0,374.2311197916667,512.0
38
+ instrumentation,phi4_mini,gsm8k,12,12,12,0,12,2,12,12,7473,7473,32,32,32,32,0,0,32,32,0,0,78.09375,121.0,224.76041666666666,2048.0
39
+ instrumentation,phi4_mini,math,12,12,12,0,12,12,12,12,7500,7500,64,64,64,64,0,0,64,64,0,0,76.03125,401.0,365.5989583333333,512.0
40
+ instrumentation,qwen25_3b,gsm8k,12,6,6,0,6,0,6,6,7473,7473,32,32,32,32,0,0,0,32,0,0,108.0,154.0,296.46875,1846.0
41
+ instrumentation,qwen25_3b,math,12,6,6,0,6,6,6,6,7500,7500,64,64,64,64,0,0,0,64,0,0,104.203125,437.0,391.5052083333333,512.0
42
+ original_100_step,qwen25_1p5b,math,6,6,6,0,600,600,600,600,7500,7500,7500,7500,7500,7500,0,0,7424,7424,1,1,119.669453125,1698.0,536.3655361938477,1024.0
43
+ original_100_step,qwen25_3b,gsm8k,6,6,6,0,600,50,600,600,7473,7473,7473,7473,7473,7473,0,0,7424,7424,1,1,103.5921875,256.0,266.59979443868,1024.0
44
+ original_40_step,qwen25_1p5b,math,9,9,9,0,360,360,360,360,7500,7500,7500,7500,7500,7500,0,0,5120,5120,0,0,120.7890625,1698.0,536.1108983357748,1024.0
45
+ original_40_step,qwen25_3b,gsm8k,9,6,6,0,240,25,240,240,7473,7473,7473,7473,7473,7473,0,0,0,5120,0,0,103.5294921875,256.0,283.0554812113444,1024.0
46
+ sampling_calibration,qwen25_1p5b,gsm8k,2,2,2,0,2,0,2,2,7473,7473,32,32,32,32,0,0,32,32,0,0,108.0,154.0,278.1640625,776.0
47
+ sampling_calibration,qwen25_1p5b,math,2,2,2,0,2,2,2,2,7500,7500,64,64,64,64,0,0,64,64,0,0,104.203125,437.0,389.1328125,512.0
48
+ sampling_calibration,qwen25_3b,gsm8k,4,4,4,0,4,1,4,4,7473,7473,32,128,32,128,0,0,32,128,0,0,107.01171875,159.0,296.9789047241211,1024.0
49
+ sampling_calibration,qwen25_3b,math,4,4,4,0,4,4,4,4,7500,7500,128,128,128,128,0,0,128,128,0,0,105.65625,776.0,446.8874969482422,1024.0
50
+ state_release,qwen25_1p5b,gsm8k,6,6,6,0,6,0,6,6,7473,7473,32,32,32,32,0,0,32,32,0,0,108.0,154.0,286.5885416666667,1225.0
51
+ state_release,qwen25_1p5b,math,6,3,3,0,3,3,3,3,7500,7500,64,64,64,64,0,0,0,64,0,0,104.203125,437.0,381.6276041666667,512.0
52
+ state_release,qwen25_3b,gsm8k,12,9,9,0,9,0,9,9,7473,7473,32,128,32,128,0,0,0,128,0,0,107.34114583333333,159.0,292.1677076551649,723.0
53
+ state_release,qwen25_3b,math,12,9,9,0,9,9,9,9,7500,7500,128,128,128,128,0,0,0,128,0,0,105.65625,776.0,449.5609368218316,1024.0
tables/review_revision/realized-workload-trials.csv ADDED
The diff for this file is too large to render. See raw diff
 
tables/review_revision/recorded-attempts.csv ADDED
The diff for this file is too large to render. See raw diff
 
tables/review_revision/release-idle-baselines.csv ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ pair_id,job_id,period,summed_idle_memory_mib,max_gpu_idle_memory_mib,baseline_record,criterion
2
+ strong-node-qwen25_1p5b-grpo-gsm8k-near_boundary-s81,5031041,1,0.0,0.0,profiles/strengthening/same_node/5031041/period1-baseline-0.csv,sum_across_two_gpus_below_2048_mib
3
+ strong-node-qwen25_1p5b-grpo-gsm8k-near_boundary-s81,5031041,2,0.0,0.0,profiles/strengthening/same_node/5031041/period2-baseline-0.csv,sum_across_two_gpus_below_2048_mib
4
+ strong-node-qwen25_1p5b-grpo-gsm8k-near_boundary-s82,5031120,1,0.0,0.0,profiles/strengthening/same_node/5031120/period1-baseline-1.csv,sum_across_two_gpus_below_2048_mib
5
+ strong-node-qwen25_1p5b-grpo-gsm8k-near_boundary-s82,5031120,2,0.0,0.0,profiles/strengthening/same_node/5031120/period2-baseline-1.csv,sum_across_two_gpus_below_2048_mib
6
+ strong-node-qwen25_1p5b-grpo-gsm8k-near_boundary-s83,5031043,1,0.0,0.0,profiles/strengthening/same_node/5031043/period1-baseline-2.csv,sum_across_two_gpus_below_2048_mib
7
+ strong-node-qwen25_1p5b-grpo-gsm8k-near_boundary-s83,5031043,2,0.0,0.0,profiles/strengthening/same_node/5031043/period2-baseline-2.csv,sum_across_two_gpus_below_2048_mib
8
+ strong-node-qwen25_1p5b-grpo-math-near_boundary-s81,5031044,1,0.0,0.0,profiles/strengthening/same_node/5031044/period1-baseline-3.csv,sum_across_two_gpus_below_2048_mib
9
+ strong-node-qwen25_1p5b-grpo-math-near_boundary-s81,5031044,2,0.0,0.0,profiles/strengthening/same_node/5031044/period2-baseline-3.csv,sum_across_two_gpus_below_2048_mib
10
+ strong-node-qwen25_1p5b-grpo-math-near_boundary-s82,5031045,1,0.0,0.0,profiles/strengthening/same_node/5031045/period1-baseline-4.csv,sum_across_two_gpus_below_2048_mib
11
+ strong-node-qwen25_1p5b-grpo-math-near_boundary-s82,5031045,2,0.0,0.0,profiles/strengthening/same_node/5031045/period2-baseline-4.csv,sum_across_two_gpus_below_2048_mib
12
+ strong-node-qwen25_1p5b-grpo-math-near_boundary-s83,5031046,1,0.0,0.0,profiles/strengthening/same_node/5031046/period1-baseline-5.csv,sum_across_two_gpus_below_2048_mib
13
+ strong-node-qwen25_1p5b-grpo-math-near_boundary-s83,5031046,2,0.0,0.0,profiles/strengthening/same_node/5031046/period2-baseline-5.csv,sum_across_two_gpus_below_2048_mib
14
+ strong-node-qwen25_3b-grpo-gsm8k-near_boundary-s81,5031091,1,0.0,0.0,profiles/strengthening/same_node/5031091/period1-baseline-12.csv,sum_across_two_gpus_below_2048_mib
15
+ strong-node-qwen25_3b-grpo-gsm8k-near_boundary-s81,5031091,2,0.0,0.0,profiles/strengthening/same_node/5031091/period2-baseline-12.csv,sum_across_two_gpus_below_2048_mib
16
+ strong-node-qwen25_3b-grpo-gsm8k-near_boundary-s82,5031092,1,0.0,0.0,profiles/strengthening/same_node/5031092/period1-baseline-13.csv,sum_across_two_gpus_below_2048_mib
17
+ strong-node-qwen25_3b-grpo-gsm8k-near_boundary-s82,5031092,2,0.0,0.0,profiles/strengthening/same_node/5031092/period2-baseline-13.csv,sum_across_two_gpus_below_2048_mib
18
+ strong-node-qwen25_3b-grpo-gsm8k-near_boundary-s83,5031094,1,0.0,0.0,profiles/strengthening/same_node/5031094/period1-baseline-14.csv,sum_across_two_gpus_below_2048_mib
19
+ strong-node-qwen25_3b-grpo-gsm8k-near_boundary-s83,5031094,2,0.0,0.0,profiles/strengthening/same_node/5031094/period2-baseline-14.csv,sum_across_two_gpus_below_2048_mib
20
+ strong-node-qwen25_3b-grpo-gsm8k-safe_margin-s81,5031101,1,0.0,0.0,profiles/strengthening/same_node/5031101/period1-baseline-15.csv,sum_across_two_gpus_below_2048_mib
21
+ strong-node-qwen25_3b-grpo-gsm8k-safe_margin-s81,5031101,2,0.0,0.0,profiles/strengthening/same_node/5031101/period2-baseline-15.csv,sum_across_two_gpus_below_2048_mib
22
+ strong-node-qwen25_3b-grpo-gsm8k-safe_margin-s82,5031103,1,0.0,0.0,profiles/strengthening/same_node/5031103/period1-baseline-16.csv,sum_across_two_gpus_below_2048_mib
23
+ strong-node-qwen25_3b-grpo-gsm8k-safe_margin-s82,5031103,2,0.0,0.0,profiles/strengthening/same_node/5031103/period2-baseline-16.csv,sum_across_two_gpus_below_2048_mib
24
+ strong-node-qwen25_3b-grpo-gsm8k-safe_margin-s83,5031106,1,0.0,0.0,profiles/strengthening/same_node/5031106/period1-baseline-17.csv,sum_across_two_gpus_below_2048_mib
25
+ strong-node-qwen25_3b-grpo-gsm8k-safe_margin-s83,5031106,2,0.0,0.0,profiles/strengthening/same_node/5031106/period2-baseline-17.csv,sum_across_two_gpus_below_2048_mib
26
+ strong-node-qwen25_3b-grpo-math-near_boundary-s81,5031124,1,0.0,0.0,profiles/strengthening/same_node/5031124/period1-baseline-18.csv,sum_across_two_gpus_below_2048_mib
27
+ strong-node-qwen25_3b-grpo-math-near_boundary-s81,5031124,2,0.0,0.0,profiles/strengthening/same_node/5031124/period2-baseline-18.csv,sum_across_two_gpus_below_2048_mib
28
+ strong-node-qwen25_3b-grpo-math-near_boundary-s82,5031126,1,0.0,0.0,profiles/strengthening/same_node/5031126/period1-baseline-19.csv,sum_across_two_gpus_below_2048_mib
29
+ strong-node-qwen25_3b-grpo-math-near_boundary-s82,5031126,2,0.0,0.0,profiles/strengthening/same_node/5031126/period2-baseline-19.csv,sum_across_two_gpus_below_2048_mib
30
+ strong-node-qwen25_3b-grpo-math-near_boundary-s83,5031131,1,0.0,0.0,profiles/strengthening/same_node/5031131/period1-baseline-20.csv,sum_across_two_gpus_below_2048_mib
31
+ strong-node-qwen25_3b-grpo-math-near_boundary-s83,5031131,2,0.0,0.0,profiles/strengthening/same_node/5031131/period2-baseline-20.csv,sum_across_two_gpus_below_2048_mib
32
+ strong-node-qwen25_3b-grpo-math-safe_margin-s81,5031603,1,0.0,0.0,profiles/strengthening/same_node/5031603/period1-baseline-21.csv,sum_across_two_gpus_below_2048_mib
33
+ strong-node-qwen25_3b-grpo-math-safe_margin-s81,5031603,2,0.0,0.0,profiles/strengthening/same_node/5031603/period2-baseline-21.csv,sum_across_two_gpus_below_2048_mib
34
+ strong-node-qwen25_3b-grpo-math-safe_margin-s82,5031148,1,0.0,0.0,profiles/strengthening/same_node/5031148/period1-baseline-22.csv,sum_across_two_gpus_below_2048_mib
35
+ strong-node-qwen25_3b-grpo-math-safe_margin-s82,5031148,2,0.0,0.0,profiles/strengthening/same_node/5031148/period2-baseline-22.csv,sum_across_two_gpus_below_2048_mib
36
+ strong-node-qwen25_3b-grpo-math-safe_margin-s83,5031149,1,0.0,0.0,profiles/strengthening/same_node/5031149/period1-baseline-23.csv,sum_across_two_gpus_below_2048_mib
37
+ strong-node-qwen25_3b-grpo-math-safe_margin-s83,5031149,2,0.0,0.0,profiles/strengthening/same_node/5031149/period2-baseline-23.csv,sum_across_two_gpus_below_2048_mib
tables/review_revision/state-release-design.csv ADDED
@@ -0,0 +1,7 @@
 
 
 
 
 
 
 
 
1
+ model_family,model,dataset,risk_tier,max_prompt_length,max_response_length,max_model_len,max_num_seqs,rollout_n,train_batch_size,train_max_samples,actor_micro_batch,rollout_logprob_micro_batch,ref_logprob_micro_batch,vllm_gpu_memory_utilization,parameter_offload,optimizer_offload,total_training_steps,save_freq,test_freq,val_before_train,source_config_id,lora_merge,release_completed,release_failure_kinds,resident_completed,resident_failure_kinds
2
+ qwen25_1p5b,Qwen/Qwen2.5-1.5B-Instruct,gsm8k,near_boundary,1024,2048,3072,64,2,32,32,16,8,8,0.7,False,False,1,-1,-1,False,grpo|qwen25_1p5b|gsm8k|2gpu|rollout2|prompt1024|response2048|batch32|1|16|8|8|0.7|False|False|True|1024|2048|3072|2|32|64,False,3,,3,
3
+ qwen25_1p5b,Qwen/Qwen2.5-1.5B-Instruct,math,near_boundary,512,512,1024,128,2,64,64,16,8,8,0.7,False,False,1,-1,-1,False,grpo|qwen25_1p5b|math|2gpu|rollout2|prompt512|response512|batch64|1|16|8|8|0.7|False|False|True|512|512|1024|2|64|128,False,3,,0,actor_update_oom
4
+ qwen25_3b,Qwen/Qwen2.5-3B-Instruct,gsm8k,safe_margin,1024,2048,3072,64,2,32,32,1,4,4,0.3,True,False,1,-1,-1,False,grpo|qwen25_3b|gsm8k|2gpu|rollout2|prompt1024|response2048|batch32|1|1|4|4|0.3|True|False|True|1024|2048|3072|2|32|64,False,3,,3,
5
+ qwen25_3b,Qwen/Qwen2.5-3B-Instruct,gsm8k,near_boundary,256,1024,1280,256,5,128,128,16,8,8,0.65,False,False,1,-1,-1,False,grpo|qwen25_3b|gsm8k|2gpu|rollout5|prompt256|response1024|batch128|1|16|8|8|0.65|False|False|True|256|1024|1280|5|128|256,False,3,,0,weight_sync_oom
6
+ qwen25_3b,Qwen/Qwen2.5-3B-Instruct,math,near_boundary,1792,1024,2816,256,5,128,128,16,8,8,0.65,False,False,1,-1,-1,False,grpo|qwen25_3b|math|2gpu|rollout5|prompt1792|response1024|batch128|1|16|8|8|0.65|False|False|True|1792|1024|2816|5|128|256,False,3,,0,weight_sync_oom
7
+ qwen25_3b,Qwen/Qwen2.5-3B-Instruct,math,safe_margin,1792,1024,2816,256,5,128,128,2,8,8,0.35,True,False,1,-1,-1,False,grpo|qwen25_3b|math|2gpu|rollout5|prompt1792|response1024|batch128|1|2|8|8|0.35|True|False|True|1792|1024|2816|5|128|256,False,3,,3,
tables/review_revision/summary.json ADDED
@@ -0,0 +1,617 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "analysis_type": "retrospective_review_driven",
3
+ "input_table_root_normalized_sha256": {
4
+ "boundary-trials.csv": "55b02090fb0c1fb96a3225235014b5e13985570a974d5f7dfb71cd2912630a1f",
5
+ "factorial-cells.csv": "8b6bf295d6fed3c98ce1ff23abcce9239a6281704d3e781b4dfecda36d8c8a37",
6
+ "factorial-inference.csv": "d9d959d8565e897bcd0f215f73a5a0dfa7bf2c9900ed0b3be0e92d67c8faeae7",
7
+ "factorial-pairs.csv": "08e05c8386411248d76b8754df7bce82c7227aa5873d702f5b61217d8838b4ce",
8
+ "factorial-phase-blocks.csv": "467e531d7f04f0984e5af4d550790838366acf3f3c4506ef620654399f896f91",
9
+ "factorial-phase-inference.csv": "9a9df549bf2c24a378a4d0ebd97d2abdb972bdf5955d080b5453c02bf1c991c7",
10
+ "factorial-phase-trials.csv": "73532896187745744bc87f034153687b4db0de4c2ec5be52b8488bebaee4f12e",
11
+ "factorial-trials.csv": "8f38cccde169a28fc11b8510e031e0ab8ec404fc87b85e9f775ddcc17ae7408f",
12
+ "historical-temporal-source-audit.csv": "11c60149f856671068831358f3ec894d958aba1ff5f5443ed5c755f7d4e128a8",
13
+ "historical-temporal-source-trials.csv": "e6badd5733c3699c53e0512e70fa53c7bb18c7328d2cb6d92e1144a986c54248",
14
+ "revision-allocation-provenance.csv": "e31a0801f41a62906a5066667d694ba544e7ed8309c3a45a29de01e71699dd14",
15
+ "revision-excluded-allocation-attempts.csv": "a42c04d96ac30471a60efd218e01e09afb7fd20fd6859ec2f1ca81892462d08d",
16
+ "revision-instrumentation-cells.csv": "cc8ba9511ef4733710082090bd03d0c5b477117fc9b69d54306313dbb79691c0",
17
+ "revision-instrumentation-inference.csv": "a51494036052ae1ea088758822ef0e0e1c1d23c6f1412e3f00cd87f3b22871cc",
18
+ "revision-instrumentation-pairs.csv": "ea852b4e35460a5e62840bcbc804c44c7fe739892a6f116e5351684ae872b08e",
19
+ "revision-instrumentation-trials.csv": "26aa324b54027214e033fc379ce59328618e1308e49286c291a843871d71428e",
20
+ "revision-temporal-cases.csv": "9418c5e274878d3b2d35bf0d514b86145f1d72528cc28b62fbc31fa7c4caa171",
21
+ "revision-temporal-pairs.csv": "8684dd8a13b0ce6d6bf3aed04ec2de51c0d5eb5aa71c5d834ab499910de349f9",
22
+ "revision-temporal-trials.csv": "466b915bb505e684c87bbf028004815ff3659cfb2b688be095b8f475576458fd",
23
+ "revision-topology-cells.csv": "bbe2416d3aebf50bc76923fd474f03dc98c02768b5638ecf42adff5a720a769b",
24
+ "revision-topology-trials.csv": "acb2cab8a61e04daa7b9ff65b211c8cc1c3d508f32de3000e1d6ac96e07b8d2d",
25
+ "same-node-inference.csv": "c31f1abd38ad455ce4f06e1a8063a4e34b338206051fda3d82508dd448022988",
26
+ "same-node-pairs.csv": "bc4ffac15897b230f73565ceff22fb8ed546ff5cb5a350dd827d6729de9e2235",
27
+ "same-node-provenance.csv": "6c14b61869c819fa28fb95f6f8b9ef62509ce0cf0f722c59d7e8548510734c3d",
28
+ "same-node-trials.csv": "65c02acefc00ccbf6442de3eb8963730d9e28e154fb495747c7914ca2435996d",
29
+ "telemetry-downsampling.csv": "f32803911e3702b811cc2f6050075f30b2d7fe7c7fddb90f06904933b3537461",
30
+ "telemetry-overhead.csv": "eed6580e9a0fe06e639c531b2994a1e3813cd3e70e4b8430cfcf1707e5e01e8b",
31
+ "telemetry-trials.csv": "a1721c7740b3e51d30670969ce45756062e8c717188b40702d73b00507831e91",
32
+ "temporal-100-inference.csv": "14ba9ecc8af5cf268bd26140adde6c0234a7f67ff61e0e8facef9832fd71cc0b",
33
+ "temporal-100-matrix.csv": "65c028828d3cd3a5a22cd42eeeaafb20ee7094f52c09d3cbab4e58b833b4f508",
34
+ "temporal-100-trials.csv": "ba45ea329414798a82307a32cdca7dbb7a09194b82c51228671facc3315f0ccd",
35
+ "temporal-40-audit.csv": "229f9dd0fae58d1ccdd81a23292cf27faac0cd02bb0c44306c02570b46cef897",
36
+ "temporal-40-summary.csv": "19ab950c596d3f48235ac215b53d0f4577797d5b4538fceffa7a2d51860653ea",
37
+ "temporal-40-trials.csv": "c462a0a01419d0ac5eae2168349ac17f1705f411f80a75d74ed125f625afdd71",
38
+ "three-model-cases.csv": "8c4ddf18c4e4734f3db05838815808ff2500115ad129a00afd584b459425b009",
39
+ "three-model-configurations.csv": "b5f4b25f358fcbc131e754b720ed2691d40cab46f9923de40fd0bf56e35fc49a",
40
+ "three-model-transfer.csv": "5546f337cb18290c485e8412bf2e77e98d7308a677011880d7b4eef9e36c4de9",
41
+ "threshold-cases.csv": "29d2614b3317e8f635cefd7d4edca2b47b7449964be352f4c6fb57bcefed577c",
42
+ "threshold-transfers.csv": "85b1d0f0006e1dcf9489c9bdd4d1992c7f9102bd8b92d98a1ef18efb09fa089e"
43
+ },
44
+ "decision_summaries": [
45
+ {
46
+ "rule": "source_model",
47
+ "donor": "granite33_2b",
48
+ "target": "phi4_mini",
49
+ "decisions": 24,
50
+ "correct_three_state": 16,
51
+ "approved": 16,
52
+ "approved_memory_failure": 0,
53
+ "approved_above_margin": 4,
54
+ "rejected_within_margin": 0,
55
+ "target_within_margin": 12,
56
+ "target_above_margin": 4,
57
+ "target_memory_failure": 8,
58
+ "within_margin_to_within_margin": 12,
59
+ "within_margin_to_above_margin": 4,
60
+ "within_margin_to_memory_failure": 0,
61
+ "above_margin_to_within_margin": 0,
62
+ "above_margin_to_above_margin": 0,
63
+ "above_margin_to_memory_failure": 4,
64
+ "memory_failure_to_within_margin": 0,
65
+ "memory_failure_to_above_margin": 0,
66
+ "memory_failure_to_memory_failure": 4
67
+ },
68
+ {
69
+ "rule": "source_model",
70
+ "donor": "granite33_2b",
71
+ "target": "qwen25_3b",
72
+ "decisions": 24,
73
+ "correct_three_state": 20,
74
+ "approved": 16,
75
+ "approved_memory_failure": 0,
76
+ "approved_above_margin": 0,
77
+ "rejected_within_margin": 0,
78
+ "target_within_margin": 16,
79
+ "target_above_margin": 0,
80
+ "target_memory_failure": 8,
81
+ "within_margin_to_within_margin": 16,
82
+ "within_margin_to_above_margin": 0,
83
+ "within_margin_to_memory_failure": 0,
84
+ "above_margin_to_within_margin": 0,
85
+ "above_margin_to_above_margin": 0,
86
+ "above_margin_to_memory_failure": 4,
87
+ "memory_failure_to_within_margin": 0,
88
+ "memory_failure_to_above_margin": 0,
89
+ "memory_failure_to_memory_failure": 4
90
+ },
91
+ {
92
+ "rule": "source_model",
93
+ "donor": "phi4_mini",
94
+ "target": "granite33_2b",
95
+ "decisions": 24,
96
+ "correct_three_state": 16,
97
+ "approved": 12,
98
+ "approved_memory_failure": 0,
99
+ "approved_above_margin": 0,
100
+ "rejected_within_margin": 4,
101
+ "target_within_margin": 16,
102
+ "target_above_margin": 4,
103
+ "target_memory_failure": 4,
104
+ "within_margin_to_within_margin": 12,
105
+ "within_margin_to_above_margin": 0,
106
+ "within_margin_to_memory_failure": 0,
107
+ "above_margin_to_within_margin": 4,
108
+ "above_margin_to_above_margin": 0,
109
+ "above_margin_to_memory_failure": 0,
110
+ "memory_failure_to_within_margin": 0,
111
+ "memory_failure_to_above_margin": 4,
112
+ "memory_failure_to_memory_failure": 4
113
+ },
114
+ {
115
+ "rule": "source_model",
116
+ "donor": "phi4_mini",
117
+ "target": "qwen25_3b",
118
+ "decisions": 24,
119
+ "correct_three_state": 20,
120
+ "approved": 12,
121
+ "approved_memory_failure": 0,
122
+ "approved_above_margin": 0,
123
+ "rejected_within_margin": 4,
124
+ "target_within_margin": 16,
125
+ "target_above_margin": 0,
126
+ "target_memory_failure": 8,
127
+ "within_margin_to_within_margin": 12,
128
+ "within_margin_to_above_margin": 0,
129
+ "within_margin_to_memory_failure": 0,
130
+ "above_margin_to_within_margin": 4,
131
+ "above_margin_to_above_margin": 0,
132
+ "above_margin_to_memory_failure": 0,
133
+ "memory_failure_to_within_margin": 0,
134
+ "memory_failure_to_above_margin": 0,
135
+ "memory_failure_to_memory_failure": 8
136
+ },
137
+ {
138
+ "rule": "source_model",
139
+ "donor": "qwen25_3b",
140
+ "target": "granite33_2b",
141
+ "decisions": 24,
142
+ "correct_three_state": 20,
143
+ "approved": 16,
144
+ "approved_memory_failure": 0,
145
+ "approved_above_margin": 0,
146
+ "rejected_within_margin": 0,
147
+ "target_within_margin": 16,
148
+ "target_above_margin": 4,
149
+ "target_memory_failure": 4,
150
+ "within_margin_to_within_margin": 16,
151
+ "within_margin_to_above_margin": 0,
152
+ "within_margin_to_memory_failure": 0,
153
+ "above_margin_to_within_margin": 0,
154
+ "above_margin_to_above_margin": 0,
155
+ "above_margin_to_memory_failure": 0,
156
+ "memory_failure_to_within_margin": 0,
157
+ "memory_failure_to_above_margin": 4,
158
+ "memory_failure_to_memory_failure": 4
159
+ },
160
+ {
161
+ "rule": "source_model",
162
+ "donor": "qwen25_3b",
163
+ "target": "phi4_mini",
164
+ "decisions": 24,
165
+ "correct_three_state": 20,
166
+ "approved": 16,
167
+ "approved_memory_failure": 0,
168
+ "approved_above_margin": 4,
169
+ "rejected_within_margin": 0,
170
+ "target_within_margin": 12,
171
+ "target_above_margin": 4,
172
+ "target_memory_failure": 8,
173
+ "within_margin_to_within_margin": 12,
174
+ "within_margin_to_above_margin": 4,
175
+ "within_margin_to_memory_failure": 0,
176
+ "above_margin_to_within_margin": 0,
177
+ "above_margin_to_above_margin": 0,
178
+ "above_margin_to_memory_failure": 0,
179
+ "memory_failure_to_within_margin": 0,
180
+ "memory_failure_to_above_margin": 0,
181
+ "memory_failure_to_memory_failure": 8
182
+ },
183
+ {
184
+ "rule": "donor_workload_lookup",
185
+ "model_family": "granite33_2b",
186
+ "donor": "code_heavy_tail",
187
+ "target": "other_three_workloads",
188
+ "included_donor_processes": 18,
189
+ "decisions": 18,
190
+ "correct_three_state": 18,
191
+ "approved": 12,
192
+ "approved_memory_failure": 0,
193
+ "approved_above_margin": 0,
194
+ "rejected_within_margin": 0,
195
+ "target_within_margin": 12,
196
+ "target_above_margin": 3,
197
+ "target_memory_failure": 3,
198
+ "within_margin_to_within_margin": 12,
199
+ "within_margin_to_above_margin": 0,
200
+ "within_margin_to_memory_failure": 0,
201
+ "above_margin_to_within_margin": 0,
202
+ "above_margin_to_above_margin": 3,
203
+ "above_margin_to_memory_failure": 0,
204
+ "memory_failure_to_within_margin": 0,
205
+ "memory_failure_to_above_margin": 0,
206
+ "memory_failure_to_memory_failure": 3
207
+ },
208
+ {
209
+ "rule": "donor_workload_lookup",
210
+ "model_family": "granite33_2b",
211
+ "donor": "code_standard",
212
+ "target": "other_three_workloads",
213
+ "included_donor_processes": 18,
214
+ "decisions": 18,
215
+ "correct_three_state": 18,
216
+ "approved": 12,
217
+ "approved_memory_failure": 0,
218
+ "approved_above_margin": 0,
219
+ "rejected_within_margin": 0,
220
+ "target_within_margin": 12,
221
+ "target_above_margin": 3,
222
+ "target_memory_failure": 3,
223
+ "within_margin_to_within_margin": 12,
224
+ "within_margin_to_above_margin": 0,
225
+ "within_margin_to_memory_failure": 0,
226
+ "above_margin_to_within_margin": 0,
227
+ "above_margin_to_above_margin": 3,
228
+ "above_margin_to_memory_failure": 0,
229
+ "memory_failure_to_within_margin": 0,
230
+ "memory_failure_to_above_margin": 0,
231
+ "memory_failure_to_memory_failure": 3
232
+ },
233
+ {
234
+ "rule": "donor_workload_lookup",
235
+ "model_family": "granite33_2b",
236
+ "donor": "gsm8k",
237
+ "target": "other_three_workloads",
238
+ "included_donor_processes": 18,
239
+ "decisions": 18,
240
+ "correct_three_state": 18,
241
+ "approved": 12,
242
+ "approved_memory_failure": 0,
243
+ "approved_above_margin": 0,
244
+ "rejected_within_margin": 0,
245
+ "target_within_margin": 12,
246
+ "target_above_margin": 3,
247
+ "target_memory_failure": 3,
248
+ "within_margin_to_within_margin": 12,
249
+ "within_margin_to_above_margin": 0,
250
+ "within_margin_to_memory_failure": 0,
251
+ "above_margin_to_within_margin": 0,
252
+ "above_margin_to_above_margin": 3,
253
+ "above_margin_to_memory_failure": 0,
254
+ "memory_failure_to_within_margin": 0,
255
+ "memory_failure_to_above_margin": 0,
256
+ "memory_failure_to_memory_failure": 3
257
+ },
258
+ {
259
+ "rule": "donor_workload_lookup",
260
+ "model_family": "granite33_2b",
261
+ "donor": "math",
262
+ "target": "other_three_workloads",
263
+ "included_donor_processes": 18,
264
+ "decisions": 18,
265
+ "correct_three_state": 18,
266
+ "approved": 12,
267
+ "approved_memory_failure": 0,
268
+ "approved_above_margin": 0,
269
+ "rejected_within_margin": 0,
270
+ "target_within_margin": 12,
271
+ "target_above_margin": 3,
272
+ "target_memory_failure": 3,
273
+ "within_margin_to_within_margin": 12,
274
+ "within_margin_to_above_margin": 0,
275
+ "within_margin_to_memory_failure": 0,
276
+ "above_margin_to_within_margin": 0,
277
+ "above_margin_to_above_margin": 3,
278
+ "above_margin_to_memory_failure": 0,
279
+ "memory_failure_to_within_margin": 0,
280
+ "memory_failure_to_above_margin": 0,
281
+ "memory_failure_to_memory_failure": 3
282
+ },
283
+ {
284
+ "rule": "donor_workload_lookup",
285
+ "model_family": "phi4_mini",
286
+ "donor": "code_heavy_tail",
287
+ "target": "other_three_workloads",
288
+ "included_donor_processes": 18,
289
+ "decisions": 18,
290
+ "correct_three_state": 18,
291
+ "approved": 9,
292
+ "approved_memory_failure": 0,
293
+ "approved_above_margin": 0,
294
+ "rejected_within_margin": 0,
295
+ "target_within_margin": 9,
296
+ "target_above_margin": 3,
297
+ "target_memory_failure": 6,
298
+ "within_margin_to_within_margin": 9,
299
+ "within_margin_to_above_margin": 0,
300
+ "within_margin_to_memory_failure": 0,
301
+ "above_margin_to_within_margin": 0,
302
+ "above_margin_to_above_margin": 3,
303
+ "above_margin_to_memory_failure": 0,
304
+ "memory_failure_to_within_margin": 0,
305
+ "memory_failure_to_above_margin": 0,
306
+ "memory_failure_to_memory_failure": 6
307
+ },
308
+ {
309
+ "rule": "donor_workload_lookup",
310
+ "model_family": "phi4_mini",
311
+ "donor": "code_standard",
312
+ "target": "other_three_workloads",
313
+ "included_donor_processes": 18,
314
+ "decisions": 18,
315
+ "correct_three_state": 18,
316
+ "approved": 9,
317
+ "approved_memory_failure": 0,
318
+ "approved_above_margin": 0,
319
+ "rejected_within_margin": 0,
320
+ "target_within_margin": 9,
321
+ "target_above_margin": 3,
322
+ "target_memory_failure": 6,
323
+ "within_margin_to_within_margin": 9,
324
+ "within_margin_to_above_margin": 0,
325
+ "within_margin_to_memory_failure": 0,
326
+ "above_margin_to_within_margin": 0,
327
+ "above_margin_to_above_margin": 3,
328
+ "above_margin_to_memory_failure": 0,
329
+ "memory_failure_to_within_margin": 0,
330
+ "memory_failure_to_above_margin": 0,
331
+ "memory_failure_to_memory_failure": 6
332
+ },
333
+ {
334
+ "rule": "donor_workload_lookup",
335
+ "model_family": "phi4_mini",
336
+ "donor": "gsm8k",
337
+ "target": "other_three_workloads",
338
+ "included_donor_processes": 18,
339
+ "decisions": 18,
340
+ "correct_three_state": 18,
341
+ "approved": 9,
342
+ "approved_memory_failure": 0,
343
+ "approved_above_margin": 0,
344
+ "rejected_within_margin": 0,
345
+ "target_within_margin": 9,
346
+ "target_above_margin": 3,
347
+ "target_memory_failure": 6,
348
+ "within_margin_to_within_margin": 9,
349
+ "within_margin_to_above_margin": 0,
350
+ "within_margin_to_memory_failure": 0,
351
+ "above_margin_to_within_margin": 0,
352
+ "above_margin_to_above_margin": 3,
353
+ "above_margin_to_memory_failure": 0,
354
+ "memory_failure_to_within_margin": 0,
355
+ "memory_failure_to_above_margin": 0,
356
+ "memory_failure_to_memory_failure": 6
357
+ },
358
+ {
359
+ "rule": "donor_workload_lookup",
360
+ "model_family": "phi4_mini",
361
+ "donor": "math",
362
+ "target": "other_three_workloads",
363
+ "included_donor_processes": 18,
364
+ "decisions": 18,
365
+ "correct_three_state": 18,
366
+ "approved": 9,
367
+ "approved_memory_failure": 0,
368
+ "approved_above_margin": 0,
369
+ "rejected_within_margin": 0,
370
+ "target_within_margin": 9,
371
+ "target_above_margin": 3,
372
+ "target_memory_failure": 6,
373
+ "within_margin_to_within_margin": 9,
374
+ "within_margin_to_above_margin": 0,
375
+ "within_margin_to_memory_failure": 0,
376
+ "above_margin_to_within_margin": 0,
377
+ "above_margin_to_above_margin": 3,
378
+ "above_margin_to_memory_failure": 0,
379
+ "memory_failure_to_within_margin": 0,
380
+ "memory_failure_to_above_margin": 0,
381
+ "memory_failure_to_memory_failure": 6
382
+ },
383
+ {
384
+ "rule": "donor_workload_lookup",
385
+ "model_family": "qwen25_3b",
386
+ "donor": "code_heavy_tail",
387
+ "target": "other_three_workloads",
388
+ "included_donor_processes": 18,
389
+ "decisions": 18,
390
+ "correct_three_state": 18,
391
+ "approved": 12,
392
+ "approved_memory_failure": 0,
393
+ "approved_above_margin": 0,
394
+ "rejected_within_margin": 0,
395
+ "target_within_margin": 12,
396
+ "target_above_margin": 0,
397
+ "target_memory_failure": 6,
398
+ "within_margin_to_within_margin": 12,
399
+ "within_margin_to_above_margin": 0,
400
+ "within_margin_to_memory_failure": 0,
401
+ "above_margin_to_within_margin": 0,
402
+ "above_margin_to_above_margin": 0,
403
+ "above_margin_to_memory_failure": 0,
404
+ "memory_failure_to_within_margin": 0,
405
+ "memory_failure_to_above_margin": 0,
406
+ "memory_failure_to_memory_failure": 6
407
+ },
408
+ {
409
+ "rule": "donor_workload_lookup",
410
+ "model_family": "qwen25_3b",
411
+ "donor": "code_standard",
412
+ "target": "other_three_workloads",
413
+ "included_donor_processes": 18,
414
+ "decisions": 18,
415
+ "correct_three_state": 18,
416
+ "approved": 12,
417
+ "approved_memory_failure": 0,
418
+ "approved_above_margin": 0,
419
+ "rejected_within_margin": 0,
420
+ "target_within_margin": 12,
421
+ "target_above_margin": 0,
422
+ "target_memory_failure": 6,
423
+ "within_margin_to_within_margin": 12,
424
+ "within_margin_to_above_margin": 0,
425
+ "within_margin_to_memory_failure": 0,
426
+ "above_margin_to_within_margin": 0,
427
+ "above_margin_to_above_margin": 0,
428
+ "above_margin_to_memory_failure": 0,
429
+ "memory_failure_to_within_margin": 0,
430
+ "memory_failure_to_above_margin": 0,
431
+ "memory_failure_to_memory_failure": 6
432
+ },
433
+ {
434
+ "rule": "donor_workload_lookup",
435
+ "model_family": "qwen25_3b",
436
+ "donor": "gsm8k",
437
+ "target": "other_three_workloads",
438
+ "included_donor_processes": 18,
439
+ "decisions": 18,
440
+ "correct_three_state": 18,
441
+ "approved": 12,
442
+ "approved_memory_failure": 0,
443
+ "approved_above_margin": 0,
444
+ "rejected_within_margin": 0,
445
+ "target_within_margin": 12,
446
+ "target_above_margin": 0,
447
+ "target_memory_failure": 6,
448
+ "within_margin_to_within_margin": 12,
449
+ "within_margin_to_above_margin": 0,
450
+ "within_margin_to_memory_failure": 0,
451
+ "above_margin_to_within_margin": 0,
452
+ "above_margin_to_above_margin": 0,
453
+ "above_margin_to_memory_failure": 0,
454
+ "memory_failure_to_within_margin": 0,
455
+ "memory_failure_to_above_margin": 0,
456
+ "memory_failure_to_memory_failure": 6
457
+ },
458
+ {
459
+ "rule": "donor_workload_lookup",
460
+ "model_family": "qwen25_3b",
461
+ "donor": "math",
462
+ "target": "other_three_workloads",
463
+ "included_donor_processes": 18,
464
+ "decisions": 18,
465
+ "correct_three_state": 18,
466
+ "approved": 12,
467
+ "approved_memory_failure": 0,
468
+ "approved_above_margin": 0,
469
+ "rejected_within_margin": 0,
470
+ "target_within_margin": 12,
471
+ "target_above_margin": 0,
472
+ "target_memory_failure": 6,
473
+ "within_margin_to_within_margin": 12,
474
+ "within_margin_to_above_margin": 0,
475
+ "within_margin_to_memory_failure": 0,
476
+ "above_margin_to_within_margin": 0,
477
+ "above_margin_to_above_margin": 0,
478
+ "above_margin_to_memory_failure": 0,
479
+ "memory_failure_to_within_margin": 0,
480
+ "memory_failure_to_above_margin": 0,
481
+ "memory_failure_to_memory_failure": 6
482
+ },
483
+ {
484
+ "rule": "gpu_count_transfer",
485
+ "donor": "two_gpu",
486
+ "target": "four_gpu",
487
+ "decisions": 18,
488
+ "correct_three_state": 10,
489
+ "approved": 10,
490
+ "approved_memory_failure": 0,
491
+ "approved_above_margin": 0,
492
+ "rejected_within_margin": 6,
493
+ "target_within_margin": 16,
494
+ "target_above_margin": 2,
495
+ "target_memory_failure": 0,
496
+ "within_margin_to_within_margin": 10,
497
+ "within_margin_to_above_margin": 0,
498
+ "within_margin_to_memory_failure": 0,
499
+ "above_margin_to_within_margin": 4,
500
+ "above_margin_to_above_margin": 0,
501
+ "above_margin_to_memory_failure": 0,
502
+ "memory_failure_to_within_margin": 2,
503
+ "memory_failure_to_above_margin": 2,
504
+ "memory_failure_to_memory_failure": 0
505
+ },
506
+ {
507
+ "rule": "gpu_count_transfer",
508
+ "donor": "four_gpu",
509
+ "target": "two_gpu",
510
+ "decisions": 18,
511
+ "correct_three_state": 10,
512
+ "approved": 16,
513
+ "approved_memory_failure": 2,
514
+ "approved_above_margin": 4,
515
+ "rejected_within_margin": 0,
516
+ "target_within_margin": 10,
517
+ "target_above_margin": 4,
518
+ "target_memory_failure": 4,
519
+ "within_margin_to_within_margin": 10,
520
+ "within_margin_to_above_margin": 4,
521
+ "within_margin_to_memory_failure": 2,
522
+ "above_margin_to_within_margin": 0,
523
+ "above_margin_to_above_margin": 0,
524
+ "above_margin_to_memory_failure": 2,
525
+ "memory_failure_to_within_margin": 0,
526
+ "memory_failure_to_above_margin": 0,
527
+ "memory_failure_to_memory_failure": 0
528
+ }
529
+ ],
530
+ "instrumentation_actor_stage_512mib_diagnostic_triggers": [],
531
+ "factorial_actor_sets_whole_run_peak": 0,
532
+ "factorial_peak_stage_patterns": {
533
+ "rollout": 73,
534
+ "idle;weight_sync": 47,
535
+ "idle": 19,
536
+ "initializing": 4,
537
+ "rollout;unknown": 1
538
+ },
539
+ "corpus_weighting_sensitivity": [
540
+ {
541
+ "estimand": "actor_main_effect_whole_run_mib",
542
+ "four_workload_mean": -30.61111111111111,
543
+ "three_corpus_equal_weight_mean": -22.59259259259259,
544
+ "minimum_workload_effect": -95.44444444444444,
545
+ "maximum_workload_effect": 0.0
546
+ },
547
+ {
548
+ "estimand": "reservation_main_effect_whole_run_mib",
549
+ "four_workload_mean": 3998.0,
550
+ "three_corpus_equal_weight_mean": 4006.4074074074074,
551
+ "minimum_workload_effect": 3881.4444444444443,
552
+ "maximum_workload_effect": 4064.1111111111113
553
+ },
554
+ {
555
+ "estimand": "interaction_whole_run_mib",
556
+ "four_workload_mean": 20.777777777777775,
557
+ "three_corpus_equal_weight_mean": 2.2222222222222197,
558
+ "minimum_workload_effect": -74.22222222222223,
559
+ "maximum_workload_effect": 192.66666666666666
560
+ },
561
+ {
562
+ "estimand": "actor_main_effect_actor_nvml_mib",
563
+ "four_workload_mean": 3980.0555555555557,
564
+ "three_corpus_equal_weight_mean": 3573.5,
565
+ "minimum_workload_effect": 2072.8888888888887,
566
+ "maximum_workload_effect": 6267.555555555556
567
+ },
568
+ {
569
+ "estimand": "reservation_main_effect_actor_nvml_mib",
570
+ "four_workload_mean": 8.83333333333334,
571
+ "three_corpus_equal_weight_mean": 41.16666666666667,
572
+ "minimum_workload_effect": -343.55555555555554,
573
+ "maximum_workload_effect": 173.22222222222223
574
+ },
575
+ {
576
+ "estimand": "interaction_actor_nvml_mib",
577
+ "four_workload_mean": 114.66666666666666,
578
+ "three_corpus_equal_weight_mean": 147.0,
579
+ "minimum_workload_effect": -34.22222222222222,
580
+ "maximum_workload_effect": 305.55555555555554
581
+ },
582
+ {
583
+ "estimand": "actor_main_effect_actor_allocated_mib",
584
+ "four_workload_mean": 3847.7573310004336,
585
+ "three_corpus_equal_weight_mean": 3450.473614728009,
586
+ "minimum_workload_effect": 2000.4676649305557,
587
+ "maximum_workload_effect": 6072.907172309027
588
+ },
589
+ {
590
+ "estimand": "reservation_main_effect_actor_allocated_mib",
591
+ "four_workload_mean": 0.6370510525173589,
592
+ "three_corpus_equal_weight_mean": 30.767424406828702,
593
+ "minimum_workload_effect": -309.62456597222223,
594
+ "maximum_workload_effect": 133.652099609375
595
+ },
596
+ {
597
+ "estimand": "interaction_actor_allocated_mib",
598
+ "four_workload_mean": 111.80247667100694,
599
+ "three_corpus_equal_weight_mean": 148.81673177083331,
600
+ "minimum_workload_effect": -40.21929253472222,
601
+ "maximum_workload_effect": 259.77088758680554
602
+ }
603
+ ],
604
+ "outputs": {
605
+ "decision-records": 396,
606
+ "decision-summaries": 20,
607
+ "instrumentation-stage-trials": 72,
608
+ "instrumentation-stage-pairs": 36,
609
+ "instrumentation-stage-cells": 12,
610
+ "factorial-peak-stages": 144,
611
+ "factorial-workload-effects": 4,
612
+ "corpus-weighting-sensitivity": 9,
613
+ "original-temporal-case-effects": 2,
614
+ "factorial-allocation-history": 144,
615
+ "release-idle-baselines": 36
616
+ }
617
+ }
upload-manifest.json CHANGED
@@ -2,37 +2,93 @@
2
  "algorithm": "sha256",
3
  "files": {
4
  "CITATION.cff": {
5
- "sha256": "5922fbd674dc24b40dc273c86b9f7803c5a425e578fe4e5eee8e80a033493c0e",
6
- "size_bytes": 487
7
  },
8
  "README.md": {
9
- "sha256": "f6daf5d18f0aa7bb198c729ace878ed0761520ea43a6459d89e4f17e8a8e0208",
10
- "size_bytes": 3789
11
  },
12
  "RIGHTS.md": {
13
  "sha256": "1030abd1dfacc5d89c8544c60bd617e1951689f6d463192b3473db649f39bd26",
14
  "size_bytes": 996
15
  },
16
  "paper/RLVRAMBench_IEEE_Access_source.zip": {
17
- "sha256": "2d28eb52f6114ba6f8e21e85cac2add3bd1ee06a6a6b1940cf2076560d70b989",
18
- "size_bytes": 1149303
19
  },
20
  "paper/RLVRAMBench_IEEE_Access_submission.pdf": {
21
- "sha256": "8ea9b321c1e94a742d7ce3304f0f455d4a65c7d402ce69a3ec7db2acc9e648a0",
22
- "size_bytes": 466017
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
23
  },
24
  "publication-requirements.txt": {
25
  "sha256": "3c86c539ab6f053981c6c58c93135f6e085434cfd4ea112424469b25f16b6a5b",
26
  "size_bytes": 458
27
  },
28
  "release.json": {
29
- "sha256": "d48c8ead76aa89870e5e9da84448a259073329bb7439c6c8ba7ae273ec0c9a84",
30
- "size_bytes": 803
 
 
 
 
 
 
 
 
31
  },
32
  "releases/RLVRAMBench-review-artifact-ieee-access-20260914-2897476.tar.zst": {
33
  "sha256": "4a735cdc33b709d330b9894cb3b923542da2c29c29d366eccdae19c781096cfd",
34
  "size_bytes": 171120507
35
  },
 
 
 
 
36
  "tables/boundary-trials.csv": {
37
  "sha256": "6b0bda9727e04be13242e8b3eca63aa8f01a4ff9122899b59d59dd2d03d0df7a",
38
  "size_bytes": 301411
@@ -73,6 +129,114 @@
73
  "sha256": "c431f178b19ca2ade3e8eb04ac5644bc5cbc195ed910749ba32289524d431ce6",
74
  "size_bytes": 20751
75
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
76
  "tables/revision-allocation-provenance.csv": {
77
  "sha256": "e31a0801f41a62906a5066667d694ba544e7ed8309c3a45a29de01e71699dd14",
78
  "size_bytes": 16192
@@ -194,8 +358,8 @@
194
  "size_bytes": 1411
195
  },
196
  "verification/VALIDATION.md": {
197
- "sha256": "af69c99de09f3dcca8e4140f79d8a1350a204239630e623c42d1169877b65017",
198
- "size_bytes": 1358
199
  },
200
  "verification/local-reproduction.json": {
201
  "sha256": "1fd391dfb571acb3d50133433605d35ea2c63e2281028092de2c114c1f69f72f",
@@ -204,8 +368,14 @@
204
  "verification/package-preparation.json": {
205
  "sha256": "bc21f03a41e9a989520260f705fa0a46abe6671630ec7ff568cbda63d83b2692",
206
  "size_bytes": 153
 
 
 
 
 
 
 
 
207
  }
208
- },
209
- "self_excluded": "upload-manifest.json",
210
- "archive_contents_scanned_separately": true
211
  }
 
2
  "algorithm": "sha256",
3
  "files": {
4
  "CITATION.cff": {
5
+ "sha256": "55a85382e8c6834ed785d81596a8170ebf7b5ca95aaccb3517a3c876d05ae662",
6
+ "size_bytes": 490
7
  },
8
  "README.md": {
9
+ "sha256": "30dd47ba32202c9e028f23963add167c116822937eb3678f5f9e2f735d613764",
10
+ "size_bytes": 4503
11
  },
12
  "RIGHTS.md": {
13
  "sha256": "1030abd1dfacc5d89c8544c60bd617e1951689f6d463192b3473db649f39bd26",
14
  "size_bytes": 996
15
  },
16
  "paper/RLVRAMBench_IEEE_Access_source.zip": {
17
+ "sha256": "cb76104ff25aeeded84335cb8f08d9d6fa13d21c203b4f4a100972800a3ab385",
18
+ "size_bytes": 1147484
19
  },
20
  "paper/RLVRAMBench_IEEE_Access_submission.pdf": {
21
+ "sha256": "0eda7af5034a3f1fd9f2a531bd13af3ea5854aa9560ae64fb40d80d5914b8bc0",
22
+ "size_bytes": 434230
23
+ },
24
+ "paper/RLVRAMBench_IEEE_Access_submission.tex": {
25
+ "sha256": "4e66ceabc3b08177eefc905284b2f4d9642a0d5b46830dc26e3f04b648e4d101",
26
+ "size_bytes": 92535
27
+ },
28
+ "paper/RLVRAMBench_supplement.pdf": {
29
+ "sha256": "f012960eb466b9f01adf1dcd481f2c588a7198ad5dcf0b9692843064e387a099",
30
+ "size_bytes": 220349
31
+ },
32
+ "paper/RLVRAMBench_supplement.tex": {
33
+ "sha256": "ba3f620bb5a349e5aee7a568bf24637c7a1b66bbca0c4cdcae24a500229537a2",
34
+ "size_bytes": 17200
35
+ },
36
+ "paper/figures/review_instrumentation_batch.pdf": {
37
+ "sha256": "c80b1e6cd4bf5fcac419a883c46c4c01d765e15a31d1fc9f65d0964356074f60",
38
+ "size_bytes": 15865
39
+ },
40
+ "paper/figures/review_instrumentation_batch.png": {
41
+ "sha256": "23b742462ff86cf810dda6413d35e5d4678bc9f6c7412dbf893f6a70e2732771",
42
+ "size_bytes": 116357
43
+ },
44
+ "paper/figures/standard_grpo_boundary.pdf": {
45
+ "sha256": "f91ca1c1596c35481c36217c02e782d7f65ccba9663d2cebab863a8dbcc31613",
46
+ "size_bytes": 15767
47
+ },
48
+ "paper/figures/standard_grpo_boundary.png": {
49
+ "sha256": "7b2d35dd05bad3848d8dd60fa8b2d6e07f3f42d131767def46401510f23eeeee",
50
+ "size_bytes": 184610
51
+ },
52
+ "paper/figures/standard_grpo_mechanisms.pdf": {
53
+ "sha256": "f4e7a6ec6ed54813960b9d6a7673a57182e72d660ed9425d6e2a0a394d230f03",
54
+ "size_bytes": 17648
55
+ },
56
+ "paper/figures/standard_grpo_mechanisms.png": {
57
+ "sha256": "f3ee3f83a769c300e9527381ed90d8bb605debed14333ac8a6e024322a59f66a",
58
+ "size_bytes": 188982
59
+ },
60
+ "paper/figures/standard_grpo_temporal.pdf": {
61
+ "sha256": "183b4bc1fc0c2dfa2b9d11b76a929202449246090f4a5a73b860de2287aaacd3",
62
+ "size_bytes": 16968
63
+ },
64
+ "paper/figures/standard_grpo_temporal.png": {
65
+ "sha256": "5760e69dcf7d59885376fe734c8f642ba448a29074805d6924c175f85a1b5d3f",
66
+ "size_bytes": 98350
67
  },
68
  "publication-requirements.txt": {
69
  "sha256": "3c86c539ab6f053981c6c58c93135f6e085434cfd4ea112424469b25f16b6a5b",
70
  "size_bytes": 458
71
  },
72
  "release.json": {
73
+ "sha256": "098977d01d65e58fa2e4ae1084e5f317e8043fe47176f184a66d351f1998c191",
74
+ "size_bytes": 1135
75
+ },
76
+ "releases/RLVRAMBench-release-critical-revision-20260914-328c1c3.json": {
77
+ "sha256": "dadcec5f1795e9c619e0bb9febe9dd97862bff1d25bb30f4f6a8714e631290f5",
78
+ "size_bytes": 1035
79
+ },
80
+ "releases/RLVRAMBench-review-artifact-critical-revision-20260914-328c1c3.tar.zst": {
81
+ "sha256": "81f9703da21b1fda72627960e477acdca5d196995787272d093788d46e6b9e4d",
82
+ "size_bytes": 174064706
83
  },
84
  "releases/RLVRAMBench-review-artifact-ieee-access-20260914-2897476.tar.zst": {
85
  "sha256": "4a735cdc33b709d330b9894cb3b923542da2c29c29d366eccdae19c781096cfd",
86
  "size_bytes": 171120507
87
  },
88
+ "releases/RLVRAMBench-submission-scientific-package-critical-revision-20260914-328c1c3.zip": {
89
+ "sha256": "89b5c3068d9e0c9f6fc2cc1faebfa25b628896e3113696688bcbe254c1e16ad2",
90
+ "size_bytes": 1783963
91
+ },
92
  "tables/boundary-trials.csv": {
93
  "sha256": "6b0bda9727e04be13242e8b3eca63aa8f01a4ff9122899b59d59dd2d03d0df7a",
94
  "size_bytes": 301411
 
129
  "sha256": "c431f178b19ca2ade3e8eb04ac5644bc5cbc195ed910749ba32289524d431ce6",
130
  "size_bytes": 20751
131
  },
132
+ "tables/review_revision/attempt-flow.csv": {
133
+ "sha256": "cd7b37adb851933b58162c84f4c9f28c5cf1e9a8e798ece3a4f15ba89ec0c40b",
134
+ "size_bytes": 636
135
+ },
136
+ "tables/review_revision/attempt-workload-summary.json": {
137
+ "sha256": "3444515799ccb50b74e406c0b13120ab2fed087dbfc06fef909ad6bb824038ee",
138
+ "size_bytes": 6821
139
+ },
140
+ "tables/review_revision/control-allocation-attempts.csv": {
141
+ "sha256": "e03ffa482718694731ea40815cea98bdcb7418db96421b02145ac52ca442f5a8",
142
+ "size_bytes": 1275
143
+ },
144
+ "tables/review_revision/control-blocks.csv": {
145
+ "sha256": "e2ae759c4da46ad49e81deb071e875333ee04349a6b3a4576566656b793d3e2d",
146
+ "size_bytes": 1592
147
+ },
148
+ "tables/review_revision/control-cases.csv": {
149
+ "sha256": "b06d01bcfca5f8ebdd816f600c55062c53580edf867e2e4010d0ab7e15730695",
150
+ "size_bytes": 2869
151
+ },
152
+ "tables/review_revision/control-process-attempts.csv": {
153
+ "sha256": "cec0864418ae7a170febb26a114c8ae2312185e18171c34a6cf26cadbb4c5a3c",
154
+ "size_bytes": 5128
155
+ },
156
+ "tables/review_revision/control-provenance.csv": {
157
+ "sha256": "d0c06b95638a35e664cefddbcad69acfbb8357ff243dd9bfae7c6d6235c2c131",
158
+ "size_bytes": 1084
159
+ },
160
+ "tables/review_revision/control-realized-work.csv": {
161
+ "sha256": "ee23d5167813bfa836174fb2df8d13fcc1052fb32cad7c02fe39dd554cc6f35a",
162
+ "size_bytes": 5165
163
+ },
164
+ "tables/review_revision/control-summary.json": {
165
+ "sha256": "efc58a673e0c4123b29144d1ca20be026eda0411061a940f4077f1fc723ec909",
166
+ "size_bytes": 6909
167
+ },
168
+ "tables/review_revision/control-trials.csv": {
169
+ "sha256": "2f22d1bb91ac2d942c14acc3b0be8d2156b4586c37fbecf7466dfbdc33157c09",
170
+ "size_bytes": 44297
171
+ },
172
+ "tables/review_revision/corpus-weighting-sensitivity.csv": {
173
+ "sha256": "3f3599f0a9e0c67d55c1f09d0c1588a21f48633a29ea1619807fa216c0601b11",
174
+ "size_bytes": 1042
175
+ },
176
+ "tables/review_revision/decision-records.csv": {
177
+ "sha256": "732553a82fb9f5c7fc196c61abb57660ed08e9603f00262521cf34d61335aa95",
178
+ "size_bytes": 32793
179
+ },
180
+ "tables/review_revision/decision-summaries.csv": {
181
+ "sha256": "23cf65e4c35692739e8c4b6c87bc46029f83e915ff86b450662a91ae12b49a80",
182
+ "size_bytes": 2431
183
+ },
184
+ "tables/review_revision/factorial-allocation-history.csv": {
185
+ "sha256": "8223d9eb6e1d56850ca762e3ba2bb47f9b8bddd73426f56e245cb0991df42527",
186
+ "size_bytes": 28334
187
+ },
188
+ "tables/review_revision/factorial-peak-stages.csv": {
189
+ "sha256": "2f415781be310abf84fb832b7d5c8f49fa8f3a80b8aa0bfd13fca5c8a3b669c9",
190
+ "size_bytes": 38110
191
+ },
192
+ "tables/review_revision/factorial-workload-effects.csv": {
193
+ "sha256": "c638f10314a073297d214fdd556caae2a56d589642561938135fe43bf1bc4a39",
194
+ "size_bytes": 1034
195
+ },
196
+ "tables/review_revision/instrumentation-stage-cells.csv": {
197
+ "sha256": "5d35d0c4e89d593921fc05eaa4281d7c8cbd6466eaf0656e47d978dbd22cc4b4",
198
+ "size_bytes": 2082
199
+ },
200
+ "tables/review_revision/instrumentation-stage-pairs.csv": {
201
+ "sha256": "05c6281ec6b357423627cd118d299ef694aa7ee882f93260d08fea88ef2915c1",
202
+ "size_bytes": 4939
203
+ },
204
+ "tables/review_revision/instrumentation-stage-trials.csv": {
205
+ "sha256": "326214f1596f44b2ca2e9d3f3e02ab0919212a235b4e64f4f617b6e7caabbf46",
206
+ "size_bytes": 15928
207
+ },
208
+ "tables/review_revision/long-run-exposure.csv": {
209
+ "sha256": "c20a9a9cddb6d626d1afa439b76faaac768df4d4472f17bc1f8da7c7e50d7bce",
210
+ "size_bytes": 1177
211
+ },
212
+ "tables/review_revision/original-temporal-case-effects.csv": {
213
+ "sha256": "29a90be89d998fe68ef2a78625de3d035f98da75c89eee8b46011a60fcd80d74",
214
+ "size_bytes": 703
215
+ },
216
+ "tables/review_revision/realized-workload-cases.csv": {
217
+ "sha256": "0a97273df30f7d0fbb1b0af36d52f669ff9b52b1db90827af8c1bdb4c72da40a",
218
+ "size_bytes": 7435
219
+ },
220
+ "tables/review_revision/realized-workload-trials.csv": {
221
+ "sha256": "63a9e3de8adf81bf9b9ab84c030e45fc0abc385ee824f41fa0957ab3e4de7e64",
222
+ "size_bytes": 108870
223
+ },
224
+ "tables/review_revision/recorded-attempts.csv": {
225
+ "sha256": "9c772dc5b4b41d9ad0c6073d7bfa31d1cfa6c4fcc446092bdc4ef7f649559dfb",
226
+ "size_bytes": 559037
227
+ },
228
+ "tables/review_revision/release-idle-baselines.csv": {
229
+ "sha256": "2be7bb95796f350f5fc100a921cdc27be1f3beda7312efe1b7781393e2352d5b",
230
+ "size_bytes": 6149
231
+ },
232
+ "tables/review_revision/state-release-design.csv": {
233
+ "sha256": "83a3b99a3d0d71303dba6edfde15d62e4a928b70e9ae9504d33bccfdb1512474",
234
+ "size_bytes": 2006
235
+ },
236
+ "tables/review_revision/summary.json": {
237
+ "sha256": "429331b6129a430f9de2b2534dc53e90f9e1fd79aaae6ed5697209f0292dd7bd",
238
+ "size_bytes": 23855
239
+ },
240
  "tables/revision-allocation-provenance.csv": {
241
  "sha256": "e31a0801f41a62906a5066667d694ba544e7ed8309c3a45a29de01e71699dd14",
242
  "size_bytes": 16192
 
358
  "size_bytes": 1411
359
  },
360
  "verification/VALIDATION.md": {
361
+ "sha256": "4f0b6b04444e7fd843bd9e09e8cad07a2e4fe934af521df0c6afacf884ffea5d",
362
+ "size_bytes": 1648
363
  },
364
  "verification/local-reproduction.json": {
365
  "sha256": "1fd391dfb571acb3d50133433605d35ea2c63e2281028092de2c114c1f69f72f",
 
368
  "verification/package-preparation.json": {
369
  "sha256": "bc21f03a41e9a989520260f705fa0a46abe6671630ec7ff568cbda63d83b2692",
370
  "size_bytes": 153
371
+ },
372
+ "verification/revised-reproduction.json": {
373
+ "sha256": "a367674ae14f0bbab157f8c9e6fede2a8e5fa7d88d2c30ad9c596e48a923be4e",
374
+ "size_bytes": 603
375
+ },
376
+ "verification/revised-source-package.json": {
377
+ "sha256": "a066598a8b5565f94d6f0dcb08bb14149d0c31b7da274e6d3bf47a0768be444b",
378
+ "size_bytes": 2144
379
  }
380
+ }
 
 
381
  }
verification/VALIDATION.md CHANGED
@@ -1,4 +1,9 @@
1
- # Release validation — 2026-09-14
 
 
 
 
 
2
 
3
  The standalone package was tested with a fresh CPython 3.11 environment and
4
  the unchanged pinned publication requirements. `pip check` passed.
 
1
+ # Initial release validation — 2026-09-14
2
+
3
+ This historical report concerns the initial `ieee-access-20260914-2897476`
4
+ snapshot and its original reproduction-code version. The revised release
5
+ has separate checks in `revised-reproduction.json` and
6
+ `revised-source-package.json`; `release.json` identifies the current archive.
7
 
8
  The standalone package was tested with a fresh CPython 3.11 environment and
9
  the unchanged pinned publication requirements. `pip check` passed.
verification/revised-reproduction.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "status": "passed",
3
+ "archive_sha256": "81f9703da21b1fda72627960e477acdca5d196995787272d093788d46e6b9e4d",
4
+ "archive_members": 19511,
5
+ "manifest_entries_verified": 19509,
6
+ "principal_processes": 588,
7
+ "supporting_source_processes": 16,
8
+ "derived_csv_json_files_matched": 40,
9
+ "figures_regenerated": 4,
10
+ "revised_csv_json_files_matched": 27,
11
+ "review_control_completed_processes": 24,
12
+ "review_control_attempted_processes": 26,
13
+ "reference_profiles_removed_from_analysis_root": true,
14
+ "original_project_and_reference_tables_hidden": true,
15
+ "network_namespace_isolation_claimed": false
16
+ }