Datasets:
license: cc-by-4.0
pretty_name: Statement Portfolio (SWE-rebench-V2, 2,094 instances)
task_categories:
- text-generation
tags:
- software-engineering
- swe-agents
- problem-statements
- swe-rebench
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files: statements.parquet
Statement Portfolio
Redrafted problem statements for 2,094 feature-implementation instances from
nebius/SWE-rebench-V2,
as used in the statement-portfolio experiments of our paper. Each instance has three
redrafts of the same required change:
variant |
Style |
|---|---|
narrative |
Explains the requested change and its intended behavior in prose |
acceptance |
Leads with the criteria that a completed implementation must meet |
scope |
States what the change should include and exclude |
Redrafts were written by a multi-agent pipeline (two drafters, cross-critique, arbiter) that reads the public pull-request record (linked issues, review discussion, PR description, commit messages). The reference patch and hidden tests were never shown to the drafting models.
Fields
| Field | Description |
|---|---|
instance_id |
SWE-rebench-V2 instance id (join key) |
repo, pr_number |
Source repository and pull request |
language |
Repository language |
base_commit |
Repository snapshot the task starts from |
variant |
narrative, acceptance, or scope |
problem_statement |
The redrafted statement, exactly as given to the agent |
generation |
How the statement was produced |
6,282 rows (2,094 instances × 3 variants).
Reconstructing executable tasks
This release contains statements only. Repository snapshot, reference patch, test patch,
FAIL_TO_PASS/PASS_TO_PASS tests, and environment configuration come from
SWE-rebench-V2 at revision 475dd5e8703bb5fb22dd3c60b5d038b019eba1e0; join on
instance_id.
License and citation
Released under CC-BY-4.0, following SWE-rebench-V2. Please cite SWE-rebench-V2 and our paper when using this data.