Skip to content

Latest commit

 

History

History
97 lines (77 loc) · 3.35 KB

data-organization.md

File metadata and controls

97 lines (77 loc) · 3.35 KB

Data organization

FeatureBench uses JSONL as the contract between data-factory stages. Each line is one complete record.

Discover run/
├── repositories.jsonl
├── run.yaml
└── .featurebench/

Build run/
├── environments.jsonl
├── failures.jsonl
├── run.yaml
└── .featurebench/

Generate run/
├── tasks.jsonl
├── failures.jsonl
├── run.yaml
└── .featurebench/

repositories.jsonl is accepted by fb build. environments.jsonl is accepted by fb generate. tasks.jsonl is the final dataset.

Repository record

Each line in repositories.jsonl contains a GitHub repository snapshot.

Field Description
discovery_run_id Discover run that created the record.
github_id, node_id Stable GitHub identifiers.
full_name, name, owner Repository identity.
clone_url, html_url, api_url Repository URLs.
default_branch Default branch observed during discovery.
description, topics Repository metadata.
stars, forks, open_issues, size_kb Repository statistics.
language, license Primary language and license.
visibility, private, archived, fork Repository status.
created_at, updated_at, pushed_at, discovered_at Timestamps.

Environment record

Each line in environments.jsonl describes one validated environment.

Field Description
environment_id Stable environment identifier.
repository Canonical repository URL.
commit Exact 40-character Git commit.
language Detected primary language.
gpu_requirement none, optional, or required.
image Immutable Docker image reference.
dockerfile Self-contained Dockerfile used to build the image.
runtime Activation, Runner, limits, and container resources.
verified_test_points Tests that passed during baseline validation.

Each verified test point contains an exact Runner id and its repository-relative path.

Task record

Each line in tasks.jsonl contains a common task envelope and an Adapter payload.

Field Description
task_id Stable task identifier.
environment_id Environment that produced the task.
image Docker image used to run the task.
runtime Runtime contract inherited from the Environment.
adapter Self-contained Adapter-specific payload.

FeatureBench payload

The built-in FeatureBench Adapter writes these fields under adapter:

Field Description
type Always featurebench.
prompt Task statement shown to the coding Agent.
gold_patch Patch that restores the masked production code.
hidden_tests_patch Patch that restores hidden F2P tests.
fail_to_pass Tests that must change from failing to passing.
pass_to_pass Tests that must remain passing.
test_timeout_seconds Timeout for one task test run.
metrics.gold_patch Gold Patch line and file counts.
metrics.tests F2P and P2P counts.
metrics.top Top-object count and Gold Patch coverage.

Run files

run.yaml records the resumable command inputs. .featurebench/ stores state, events, diagnostics, invocation snapshots, and optional audit artifacts. These files support operation and debugging; they are not stage inputs.