Core Concepts

Curator has the following core concepts.

PR pool

A PR pool is a language-specific input file containing GitHub pull requests in the form:

owner/repo:pr-123

The block stores PR pools under artifacts/collected_prs/. The main language scripts pass these files to swegen create through --input-ids-file. Collection and generation are separate stages. The collector may append to a pool over time, while swegen create parses the file once at startup; wait for collection to finish when all newly collected PRs must be included. The production create scripts use the fixed artifacts/collected_prs/ paths rather than pr_collection.output_dir.

Task skeleton

A task skeleton is the Harbor task directory that Curator builds from a PR. It contains the problem instruction, Docker environment, bug patch, solution patch, and verification tests.

The stable structure is:

task_id/
├── task.toml
├── instruction.md
├── environment/
│   ├── Dockerfile
│   └── bug.patch
├── solution/
│   ├── fix.patch
│   └── solve.sh
└── tests/
    └── test.sh

The task ID is derived from the repository and PR number, for example owner__repo-123.

NOP and Oracle validation

Curator does not expose every generated skeleton downstream. It validates a candidate task with two checks:

• NOP - the unmodified buggy environment should fail the task test.
• Oracle - applying the ground-truth solution should pass the task test.

Only tasks that pass validation are considered verified.

Verified task manifest

Each language output directory has a manifest:

artifacts/swe_tasks/<lang>-cc/verifiable_tasks.txt

This file is the authoritative downstream contract. A task may exist on disk because it is in progress, failed, or partially generated, but it is only safe for tracer or other consumers when its task ID appears in verifiable_tasks.txt.

Batch state

Long runs resume through per-output batch state:

artifacts/swe_tasks/<lang>-cc/.swegen-create-batch/<hash>.json

The hash is based on the resolved absolute path of the input PR file. That means moving a restored run to a different clone path requires relocating or regenerating the batch state filename before continuing.

Batch state records each PR case, attempts, status, errors, model fingerprint, elapsed time, and selected task ID.

Language outputs

Curator uses one output directory per language:

Difficulty scoring

Curator can add static difficulty metadata to task.toml. The score is derived from the patch, tests, instruction, and file scope. The resulting fields are used for dataset analysis and sampling:

• difficulty_score
• difficulty_label
• difficulty
• category
• tags