VmaxRL/SWEUniverse-SizeMatched-45-ReliableTests-20260522

SWEUniverse Size-Matched 45 Repo Reliable Tests

This dataset contains the reliable-test universe rows collected for the 45 repositories in selected_repos_1000_size_matched.tsv.

Contents

• data/reliable_tests.jsonl: one strict-stability reliable-test row per completed repo at the 600s timeout bucket. Rows include actual stable_passing, stable_failing, excluded_tests, raw_log_refs, parser metadata, command, commit, and replica counts.
• data/snapshot_observations.jsonl: combined per-replica snapshot observations for completed repos. Each row records repo/commit/command/parser identity, run id, elapsed seconds, parsed passing/failing test names, timeout/exit status, parse errors, and raw-log references.
• metadata/raw_experiments.jsonl: raw experiment rows showing the exact command, parser, framework, repo, commit, architecture, image metadata, timeout, and replica policy used per completed repo.
• metadata/repo_summary.jsonl: compact per-repo status/count summary.
• metadata/failed_repos.json: repos that failed collection.
• metadata/selected_repos_1000_size_matched.tsv: the requested 45-repo selection/distribution file.
• metadata/swesmith_test_universe_overrides.jsonl: repository command/parser overrides available to the collector at runtime.
• metadata/target_pool/: copied target-pool manifest files used to resolve repo commits/images.
• metadata/RUN.md: operator run note.
• metadata/manager.log: wave manager log.
• metadata/run_summary.json: machine-readable run summary.

Run Summary

• VM: codex-bugpilot-runner-20260427-210206
• GCP zone: us-west1-b
• Machine type: n2-standard-128 (128 vCPU)
• Disk: 300GB boot disk, observed as 296G filesystem
• Worktree: /home/augustine_vmax_ai_com/codex-runs/pr1303-validation
• Runtime git commit: 4354af2ac2676cb19dab88a50735c19360215d44
• Reproducible PR commit with no Docker storage flag: 9298715405d5c4f8b43cf2c5a29ec27f0718d040
• Latest PR branch commit at packaging time: 29a537993
• Target-pool bundle: /home/augustine_vmax_ai_com/codex-runs/swe-universe-indist-pr1868-128c/projects/swe-turing/defaults/runtime/datasets/swesmith/canonical-swesmith-target-pool-1000-size-matched
• Target-pool manifest sha256: 1dd1b6e599c73b0a266e9a5873d75951dd97f828de8bb1600a09191dd7cc8f7c
• Selected-repos TSV sha256: d5977b8d100240155e385775cc234f703b1c0ed9917fdb38b704d25090f0917f
• Override JSONL sha256: 6487b380588db50c5962990c26191518e61261b72e87712f71b145958c523323

The runtime worktree started from 4354af2ac2676cb19dab88a50735c19360215d44 and used the no---storage-opt Docker invocation that was later committed on the PR branch as 9298715405d5c4f8b43cf2c5a29ec27f0718d040. The unused storage_size field difference is not behaviorally relevant; Docker was run with CPU/memory limits and no writable-layer storage cap.

Command

Each repo was launched through the wave manager with:

WAVE_SIZE=12 nohup .venv/bin/python /home/augustine_vmax_ai_com/runs/size-matched-45-limited-nostorage-allrepos-20260522T074643Z/run_waves.py /home/augustine_vmax_ai_com/runs/size-matched-45-limited-nostorage-allrepos-20260522T074643Z > /home/augustine_vmax_ai_com/runs/size-matched-45-limited-nostorage-allrepos-20260522T074643Z/nohup.out 2>&1 &

Each per-repo job ran:

.venv/bin/python projects/swe-turing/scripts/collect_swe_test_universes.py \
  --target-pool-bundle /home/augustine_vmax_ai_com/codex-runs/swe-universe-indist-pr1868-128c/projects/swe-turing/defaults/runtime/datasets/swesmith/canonical-swesmith-target-pool-1000-size-matched \
  --output-root "$OUT/data" \
  --dataset-id "swe-universe-size-matched-${NAME}-limited-20260522" \
  --replicas 10 \
  --max-timeout-sec 600 \
  --snapshot-interval-sec 15 \
  --timeout-buckets-sec 600 \
  --max-concurrency 10 \
  --backend docker \
  --repo-at-commit-key "$KEY" \
  --overwrite

Resource Constraints

Docker collection backend resource constraints:

• --cpus 1
• --memory 4g
• no --storage-opt writable-layer cap

The storage cap was intentionally omitted because this VM runs Docker overlay2 on ext4; Docker rejects --storage-opt size=10G unless the daemon uses a compatible storage backend such as overlay2 on XFS with project quotas.

Collection policy:

• replicas: 10
• timeout: 600 seconds
• timeout bucket: 600 seconds only
• snapshot interval: 15 seconds
• per-repo replica concurrency: 10
• repo wave size: 12

Commands, Parsers, and Overrides

The exact effective command/parser/framework for each completed repo is recorded in metadata/raw_experiments.jsonl and repeated in each reliable row in data/reliable_tests.jsonl via test_command, test_framework, and test_parser.

The collector loaded repository-specific command/parser overrides from metadata/swesmith_test_universe_overrides.jsonl at the runtime commit above. The target-pool manifest used to resolve repo commits/images is copied under metadata/target_pool/.

Results

• Completed reliable rows/repos: 43
• Snapshot observation rows: 9036
• Failed repos: iterative__dvc rc=1, conan-io__conan rc=1
• Empty stable/excluded test sets: modin-project__modin, project-monai__monai, python__mypy

Rows with empty stable/excluded sets have raw log refs and parsed replica counts but no stable or excluded test names; inspect raw logs/parser behavior before treating them as useful reliable universes.