DeNovoSWE: Whole-Repository Generation Dataset

DeNovoSWE is a large-scale dataset for whole-repository generation introduced by AweAI-Team. It contains 4,818 high-quality instances derived from 11,000 filtered trajectories. The dataset was last updated on June 14, 2026.

Use Cases

• Train models for whole-repository generation based on the dataset's high-quality instances.
• Benchmark AI performance on long-horizon software engineering tasks based on the described trajectories.
• Study code generation scaling using the filtered trajectory data mentioned in the description.

Strengths

• Contains 4,818 high-quality instances for training.
• Derived from 11,000 filtered trajectories sourced from a larger pool of 34,000 DeepSeek-v4-High trajectories.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

AweAI-Team

Collection Method

Filtered from DeepSeek-v4-High trajectories.

Freshness

Last updated 2026-06-14 08:07:10.