Datasets:

jumplander
/

JL-AgentBehavior-10K

All v1 records are generated through deterministic scenario composition using the included generator. The release does not ingest private user conversations, production telemetry, real repositories, GitHub issues, pull requests, or proprietary source code.

Repository names, commits, paths, symbols, tool results, and task scenarios are synthetic.

Personal and sensitive information

The generator is not designed to create personal data. The dataset does not intentionally contain names, email addresses, credentials, API keys, production endpoints, or private repository content.

Because all records are generated, users should still run their own secret and PII scanners after any downstream merge with external data.

Annotation

Version 1 does not contain human annotations. Behavioral labels are generated from a fixed specification. human_reviewed is false for every record.

Verification

Completed:

schema and required-field checks;
exact count and quota checks;
ID uniqueness;
fingerprint uniqueness and integrity;
split/file consistency;
repository-family isolation.

Not completed:

repository materialization;
code execution;
patch applicability;
test execution;
human preference validation;
semantic near-duplicate audit;
independent benchmark calibration.

Representation

The dataset overrepresents:

Python and TypeScript;
search/read/patch/test agent loops;
service and developer-tool scenarios;
English software-engineering language;
bounded changes with explicit constraints.

It underrepresents:

mobile and embedded systems;
frontend visual work;
data science notebooks;
distributed production incidents;
database migration execution;
hardware-specific debugging;
non-English tasks;
long-horizon multi-day engineering work.

Environmental and social impact

The dataset can be regenerated using the included standard-library Python script without model inference. Downstream model training may require substantial compute; consumers should report training hardware and energy-relevant configuration when practical.

Coding agents can increase developer productivity but can also generate insecure changes, overreach authorization, or create false confidence. Runtime sandboxing, permission controls, auditing, and independent verification remain necessary.

Maintenance

Patch releases may correct documentation or validation defects. Any record-semantic change requires at least a minor version update. Future executed or reviewed tiers must preserve per-record evidence metadata.

Contact and reporting

Report dataset defects through the public JumpLander project channels:

A useful report includes dataset version, record ID, fingerprint, failure category, and a minimal explanation.