Datasets:

FWeindel
/

validated-tasks

Exploration: find required names and calling conventions in supplied context. Specification Adherence: write the requested continuation in the required output file.

Intended benchmark transfer: Python subset: HumanEval+, MBPP+ and DS-1000 mini-200. Java/Python/TypeScript: repository exploration for SWE-bench Verified random-100. Labels follow CLI-Universe.

Per-stage breakdown

Before validation, 1,043 tasks were archived for missing context identifiers and 10 for answers with no scored identifiers. Results are from the historical validation.

1. Build and setup: 6,710 changed, 0 archived. Changed 6,710 tasks to pin verifier dependencies and support the Apptainer runtime.
2. Reference solution (reward 1): 6,710 changed, 0 archived. Added reference solutions to 6,710 tasks.
3. No-op baseline (reward 0): 6,710 changed, 0 archived. Changed scoring in 6,710 tasks to reject incomplete prefix matches and normalize comments and whitespace.
4. Rule-based checks: 6,710 changed, 0 archived. Changed 6,710 tasks to clarify context guidance and restore cross-file excerpts with answer leakage removed. No kept tasks have relative-path warnings. Initial flagged and automatically repaired task counts are missing.

Git-history and timestamp leak checks were not run.

Per-task resource requirements

Kept tasks	CPUs per task	Memory per task
6,710	Missing	Missing

Missing values are not declared in the published task configurations.

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; 48 CPU cores, 128 GiB RAM per job, 16 concurrent tasks, no GPUs.

Stage	Measured kept tasks	Total stage time	Mean per task	Median per task
1. Build and setup	6,710	Missing	16.50 s	13.36 s
2. Reference solution	6,710	Missing	14.39 s	11.20 s
3. No-op baseline	6,710	Missing	15.35 s	11.50 s
4. Rule-based checks	6,710	Missing	3.36 s	0.54 s