sb-cli 0.1.3

Submit predictions to the SWE-bench API and manage your runs

SWE-bench CLI

A command-line interface for interacting with the SWE-bench API. Use this tool to submit predictions, manage runs, and retrieve evaluation reports.

Read the full documentation here. For submission guidelines, see here.

Installation

pip install sb-cli

Authentication

Before using the CLI, you'll need to get an API key:

1. Generate an API key:

sb-cli gen-api-key [e-mail]

2. Set your API key as an environment variable - and store it somewhere safe!

export SWEBENCH_API_KEY=your_api_key
# or add export SWEBENCH_API_KEY=your_api_key to your .*rc file

3. You'll receive an email with a verification code. Verify your API key:

sb-cli verify-api-key YOUR_VERIFICATION_CODE

Subsets and Splits

SWE-bench has different subsets and splits available:

Subsets

• swe-bench-m: The main dataset
• swe-bench_lite: A smaller subset for testing and development
• swe-bench_verified: 500 verified problems from SWE-bench Learn more

Splits

• dev: Development/validation split
• test: Test split (currently only available for swe-bench_lite and swe-bench_verified)

You'll need to specify both a subset and split for most commands.

Usage

Submit Predictions

Submit your model's predictions to SWE-bench:

sb-cli submit swe-bench-m test \
    --predictions_path predictions.json \
    --run_id my_run_id

Options:

• --run_id: ID of the run to submit predictions for (optional, defaults to the name of the parent directory of the predictions file)
• --instance_ids: Comma-separated list of specific instance IDs to submit (optional)
• --output_dir: Directory to save report files (default: sb-cli-reports)
• --overwrite: Overwrite existing report (default: 0)
• --gen_report: Generate a report after evaluation is complete (default: 1)

Get Report

Retrieve evaluation results for a specific run:

sb-cli get-report swe-bench-m dev my_run_id -o ./reports

List Runs

View all your existing run IDs for a specific subset and split:

sb-cli list-runs swe-bench-m dev

Predictions File Format

Your predictions file should be a JSON file in one of these formats:

{
    "instance_id_1": {
        "model_patch": "...",
        "model_name_or_path": "..."
    },
    "instance_id_2": {
        "model_patch": "...",
        "model_name_or_path": "..."
    }
}

Or as a list:

[
    {
        "instance_id": "instance_id_1",
        "model_patch": "...",
        "model_name_or_path": "..."
    },
    {
        "instance_id": "instance_id_2",
        "model_patch": "...",
        "model_name_or_path": "..."
    }
]

Submitting to the Multimodal Leaderboard

To submit your system to the SWE-bench Multimodal leaderboard:

1. Submit your predictions for the swe-bench-m / test split using the CLI
2. Fork the experiments repository
3. Add your submission files under experiments/multimodal/YOUR_MODEL_NAME/
4. Create a PR with your submission

See the detailed guide in our submission documentation.

Note: Check your test split quota using sb-cli quota swe-bench-m test before submitting.