Datasets:
SETA-Env
Dataset stats
- Original tasks: 4,567 (1,414 excluded during upstream conversion; 3,153 audited here)
- Kept: 3,107
- Changed: 3,042
- Archived: 46
- Images: 11 images before, 11 images now.
Dataset characteristics
Domains
Data Science and Machine Learning: data cleaning, analysis, scikit-learn models and reports. Debugging: Python/pandas repairs. System Administration: shell and system configuration. Data Processing: file inventories, policy checks and structured outputs.
Capabilities
Specification Adherence: exact paths, schemas, metrics and constraints. Exploration: inspect data, policies and broken scripts. Error Recovery: reproduce failures and verify repairs. Labels follow CLI-Universe.
Intended benchmark transfer: Terminal-Bench 2.0, OpenThoughts-TBLite 2.0, DS-1000 mini-200 and BixBench CLI corrections @ Pi, through terminal workflows, Python data analysis and executable verification.
Per-stage breakdown
Includes historical runs and a fresh build/reference/no-op rerun of 3 restored tasks.
- 1. Build and setup: 3,025 changed, 24 archived. Changed 3,025 tasks with overlapping repairs to 215 environments, 72 setups, 3,021 verifiers and 35 configurations. Repairs covered runtime compatibility, dependencies and resource/time limits; old/new limit counts are missing. Archived 19 ownership mismatches, 2 read-only NVIDIA-tool conflicts and 3 slow R setups.
- 2. Reference solution (reward 1): 14 changed, 1 archived. Changed 14 reference solutions: 9 package pins and 5 script repairs covering CLI use, strict error handling, timing and signals. Archived 1 slow desktop setup.
- 3. No-op baseline (reward 0): 32 changed, 2 archived. Changed 31 setup timeouts and 1 missing-backup-script check; old/new timeouts are missing. Archived 2 slow R setups.
- 4. Rule-based checks: 356 changed, 19 archived. Clarified 25 instructions. Flagged 415 kept tasks for relative paths. The automatic resolver changed 333 tasks; 114 still have warnings. The warning-cleared count is missing. Inspected a random sample of 50, and all warnings were harmless. Archived 19 tasks requiring undisclosed deliverable paths.
Git-history and timestamp leak checks were not run.
Per-task resource requirements
| Kept tasks | CPUs per task | Memory per task |
|---|---|---|
| 3,103 | 1 | 2 GiB |
| 1 | 1 | 4 GiB |
| 1 | 1 | 6 GiB |
| 2 | 1 | 8 GiB |
Runtime and validation time
- Versions: Apptainer 1.5.4, Python 3.12.14, Harbor 0.8.1.
- Implementation: Marin Harbor
7faf878c14b6+ our runtime patches. - Resources: Helma; 32 CPU cores/job, up to 16 concurrent runtime tasks.
- Validation job memory: Missing.
| Stage | Measured tasks | Total stage time | Mean per task | Median per task |
|---|---|---|---|---|
| 1. Build and setup | 3,130 | Missing | Missing | 28.40 s |
| 2. Reference solution | 3,130 | Missing | Missing | 35.60 s |
| 3. No-op baseline | 3,130 | Missing | Missing | 38.40 s |
| 4. Rule-based checks | 3,130 | Missing | Missing | 0.70 s |
Measurements cover the original 3,130-task run, before later repairs and exclusions.