CoolFace
Datasetpublic

foreverCuSO4/workspace-data-archive

workspace-data-archive Data-artifact archive of projects under /mnt/nvme0/zhujiayi/workspace/ on the Rosetta host, packaged 2026-09-08 before the machine was wiped. Code lives on GitHub; this repo preserves the non-reproducible data products (task sets, agent trajectories, annotations, checkpoints, evaluation results). Related: foreverCuSO4/transfer-bench-Rosetta-runs-archive (self-play run archive), foreverCuSO4/transfer-bench-intents-v2-packed (3619-task dataset). Per-archive… See the full description on the dataset page: https://huggingface.co/datasets/foreverCuSO4/workspace-data-archive.

sourceHugging Facemitupdated 17d agoView on Hugging Face
0likes145downloads
Dataset Card

workspace-data-archive

Data-artifact archive of projects under /mnt/nvme0/zhujiayi/workspace/ on the Rosetta host, packaged 2026-09-08 before the machine was wiped. Code lives on GitHub; this repo preserves the non-reproducible data products (task sets, agent trajectories, annotations, checkpoints, evaluation results). Related: foreverCuSO4/transfer-bench-Rosetta-runs-archive (self-play run archive), foreverCuSO4/transfer-bench-intents-v2-packed (3619-task dataset).

Per-archive details: see the *manifest*.json files inside each project directory.

Contents

task-generation-agent/ — G4 migration-task production line (Tranfer-Tasksmith)

tarballcontent
accepted-tasks-20260706-211153-interactive-produce.tar.gz112 qualified G4 Harbor tasks + manifest (the accepted taskset, was not under git)
accepted-final-snapshots-20260702-20260705.tar.gz86 accepted final task snapshots extracted via batch-state.json (35+51)
runs-20260702-0705-0706-smalls.tar.gztrace/tasksmith/analyzer/verifier/out-minus-smoke + batch plans of the three produce batches
runs-20260702-0705-0706-solver-trajectories.tar.gzsolver trajectory.json/transcript.md/submission.diff net files
runs-20260627_packed.tar.gzJune g4-taskgen experiment archive
runs-remaining-small-runs.tar.gzall other small runs (0701, real-validation, canary, qualification-v2, g4-fdtd, fdtd-solver, reference-store…)

Dropped: workspaces/, snapshots/ intermediates, admission smoke/ (public-repo copies).

accel-trans-bench-private/ — Rosetta predecessor bench repo

  • —selfplay-batch_runs-filtered.tar.gz — 07-22/23 self-play batches: opencode.db trajectories + verdicts (pre-dates the Rosetta runs-archive, no overlap)
  • —selfplay-runs-filtered.tar.gz, batch-runs-notasksmith.tar.gz, jobs-all.tar.gz (early harbor trial artifacts), trouble-shooting-2026071x-runs.tar.gz

Dropped: workspace/pristine/tmp, tasksmith-loop-artifacts (194G public-repo snapshots).

Jittor-7/ — 计图赛道一 (dynamic recommendation)

  • —c1-data-downloads (contest raw data), c2-final-submission-cascade (final entry, val 0.6037), c3-submission-zips (10 real submissions), c4-hstu-weights (2 HSTU best.pt), c5-metrics-configs (all-experiments metrics/config harvest)

Dropped: ~255G regenerable feature caches / intermediate outputs.

migration-pipelines/ — candidate discovery pipelines

  • —jl-mtp-run-20260828-152257 — 799 g4_review cards + OpenCode agent traces + scoring CSVs + raw GitHub search frontiers (jl pipeline)
  • —jl-mtp-run-20260827-182944-unique20 — the 20 repos unique to the 0827 run
  • —mtp-run-20260617-003331-github-goal-10000 — d-review jsonl (3375 repos), candidate_cards, pipeline buffers, raw/interim/processed data (origin of the 512 taskset selection)

Dropped: 290G re-clonable public repo checkouts.

misc-robotics-ascend/ — robotics & Ascend experiments

tarballcontent
e1-dart-annotation-tool_*5251 dart-armor keypoint annotations (non-empty labels+frames) + raw capture videos + self-trained kpts models (best_0207.pt, yolov10n-kpts-latest.onnx)
e2-antidart_detection_traditional_*3 YOLO datasets (0616/0626/mixed-0205), raw videos, Downloads.zip (unique far_dataset), 5 self-trained CNN classifiers + eval metrics
e3-2026-vision-contest_*297 RL self-play .sdw checkpoints + training logs
e4-Gimble_crtl_*43 gimbal log packages (unique source) + built datasets + LSTM checkpoints/metrics + NPU HBM monitoring
e5-dLLM-cache-Ascend_*lm_eval results + perf anchors (Ascend port of dLLM-cache)
e6-Domino_Ascend_*domino speculative-decoding bench outputs
e7-transfer-bench-Rosetta-cpp_*python→cpp track intents/demo run traces
claude-code-sandbox/*.tar.gzE2B sandbox template image exports (x86_64+arm64, sha256 in manifest)

Notes

  • —Everything regenerable (public clones, pip/venv, feature caches, execution scratch) was deliberately dropped.
  • —Secrets (auth.json, tokens) are intentionally excluded everywhere.
  • —Code rescue before wipe: foreverCuSO4/dart_annotation_tool and foreverCuSO4/Gimble_LSTM were pushed to GitHub for the first time as part of this archival.

Split archives

Tarballs >1.5G were split for upload robustness: <name>.tar.gz.part-00, .part-01, ... Reassemble before extraction: cat '<name>.tar.gz.part-'* > '<name>.tar.gz' (parts are split -b 1500m, byte-exact).