Long-horizon software engineering benchmarks are rare but crucial for testing agent persistence, planning, and multi-step coding workflows. This dataset packages a compact, evaluation-focused collection of English code snippets and structured metadata in Parquet so you can iterate on agent behaviours without heavy engineering overhead.
What Sets It Apart
- Compact, evaluation-first scope: fewer than 1,000 examples makes quick experiments and CI-friendly benchmarks feasible rather than full-scale model training.
- Parquet + tabular metadata: ready-to-load with pandas/polars and huggingface datasets for deterministic, reproducible evaluation pipelines.
- Targeted for coding agents and software-engineering tasks: examples and tags oriented toward long-horizon interactions, benchmark YAML evaluation, and agent-in-the-loop workflows.
- Lightweight contributor footprint: small downloads and likes indicate a niche, curated dataset useful for early-stage research and benchmarking.
Who It's For — and Tradeoffs
Great fit if you need a small, easy-to-load benchmark to iterate on agent planning, multi-step code edits, or evaluation harnesses without managing large corpora. Look elsewhere if you require large-scale pretraining data, multilingual code coverage, or a clearly licensed corpus—this dataset is tiny by design and currently has no license declared on the card, which limits commercial or redistributable use without clarification.
