docs/data-format.md

Data Format

The default configuration contains the final 171 task rows in original release order and uses the test split. A JSONL counterpart is included for use without Hugging Face libraries.

Task rows

The Arrow/Parquet tables use fixed column types. Core fields are instance_id, original_instance_id, ordinal, language, task, problem_statement, primary repository/PR metadata, repository_count, repositories, gold_patches, test_patches, image_ref, platform.

The repositories list records each provided repository's key, URL, role, base and fixed commit, workspace path, and container path. Full source-construction, licensing, and evaluator metadata remain in the exact task files.

Statements are also available at statements/<instance_id>.md; those files and the problem_statement column contain identical text. PASS_TO_PASS_count and FAIL_TO_PASS_count count selected targets at their recorded granularity. The complete identifier lists remain in the two uppercase columns, placed after the compact metadata columns.

evaluation_support is candidate_api_detected when the entry-point inventory identifies an explicit candidate interface by static inspection, and instance_specific otherwise. This field does not claim a new runtime validation. See evaluation entry points.

Patch records contain repository, release-relative path, sha256, encoding, and content. Text is decoded as UTF-8 when valid; other bytes use base64. The hash always refers to the original patch bytes. These are evaluator materials, not agent context.

task_path, evaluation_contract_path, and environment_path resolve relative to the downloaded snapshot root. original_task_sha256 records the source task hash. data/task-bindings.jsonl records both the exported task/contract hashes and their source hashes; use its task_sha256 and contract_sha256 fields to verify the exported files. source_release_commit pins the source material release. Full file-based records in data/instances.jsonl retain their original schema and task-relative artifact paths.

Original bases and provided source projections

Task base commits retain their original provenance. Some delivered images use explicitly recorded source projections to remove pre-existing answers or address genuine compatibility gaps. Their image/build metadata identifies the delivered sources and effective bindings. The raw task base should not be silently substituted for the provided image source tree.

Reference revisions

gold_patches contains the references selected by the published task records. Approved corrections can be recorded separately in an environment's reference-revision metadata or evaluator support. Preserve the original Gold and read that task's effective reference policy when reproducing a reference endpoint. No reference patch is automatically applied to candidate mode.

references/index.json lists the separately approved Next89111 r3, Higress4430 r5, UV20804 r1 and PNPM12409 PATH revisions. The Higress full patch applies to the documented projected initial source tree. For UV, use the complete effective provider patch or the original provider Gold followed by the omission supplement; do not apply both alternatives. PNPM's existing evaluator derives its authorized revision deterministically from the preserved original artifacts. These files do not change the canonical task Gold.

Gold/test artifacts and reference versions retain their source bindings. The source and exported metadata hashes are recorded separately.

Image and release metadata

• data/reusable-agent-images.json records the 171 image-descriptor hash bindings. - data/images.jsonl provides a uniform instance-to-image table for the helper CLI. - release.json describes this HF release, source commit, original logo hash, and validation scope. - CHECKSUMS.sha256 covers all released files except itself. - data/file-modes.json records file permissions for reconstruction after HTTP downloads.

Each image descriptor identifies the published registry image and its source and evaluation metadata. Use image_ref for the immutable image; auxiliary package hashes bind the corresponding evaluator files.

Language and task annotations

In Parquet and load_dataset, language and task use Hugging Face ClassLabel features with 13 and 8 categories. Python rows store category IDs; use dataset.features[column].int2str(value) for the category name. The JSONL and TSV files retain readable names. The complete name lists are in data/categorical-features.json.

language follows section 3.1 of the 171-task statistics report: each instance has one language category, including Multi-language. task follows its section 5 eight-group display taxonomy. The three raw repair intents are grouped as Bug fixing. Labels are joined through the original instance ID, never through historical or subset row ordinals. Source hashes and exact distribution counts are in data/annotation-provenance.json. Original task files retain their construction-time metadata for provenance; these two display columns replace category and domain only in the public task table.

P2P and F2P target identifiers

The two uppercase columns use SWE-bench-style JSON-encoded string lists. They record the retained task selection, not raw execution results. Each identifier preserves its repository/mode namespace when present. If an evaluator publishes task-level IDs without repository attribution, the shared namespace is used rather than inventing an owner.

Where complete per-case identities are not available in retained records, the existing released frozen script entry is listed as a script_contract target, at its recorded granularity. Native contracts with an exact named selector retain that selector as such. data/test-target-metadata.jsonl identifies these granularities, actual artifact paths and hashes, and any accepted initial unexecuted records. A script entry can execute many underlying assertions; list lengths must not be reported as native-body execution counts. Compile/import failure or causal blocking is not converted into an executed assertion failure. Existing assertions, commands, patches, and image identities are unchanged, the lists preserve the recorded evaluation selection.

Optional image reconstruction

Manifests under an environment's rebuild/ directory describe files for reconstructing the image. They are not prerequisites for pulling and opening the published image. Evaluation contracts, package manifests and files explicitly required by an evaluator remain evaluation inputs.