agentds-bench 1.3.0

Python client for AgentDS-Bench: A streamlined benchmarking platform for evaluating AI agent capabilities in data science tasks

AgentDS Python Client

The official Python client for AgentDS-Bench, a comprehensive benchmarking platform for evaluating AI agent capabilities in data science tasks.

Features

• Seamless Authentication: Multiple authentication methods with persistent credential storage
• Task Management: Retrieve, validate, and submit responses to benchmark tasks
• Comprehensive API: Full coverage of the AgentDS-Bench platform capabilities
• Type Safety: Complete type annotations for enhanced development experience
• Professional Documentation: Extensive documentation and examples

CLI Usage

After installing, a CLI agentds is available:

# Authenticate and save credentials
agentds auth --api-key <API_KEY> --team-name <TEAM_NAME>

# Start competition
agentds start

# List domains and view status
agentds domains
agentds status

# Submit predictions
agentds submit --domain Insurance --task 1 --file predictions.csv --format "ID,Prediction" --expected-rows 1000 --numeric --prob-range 0 1

# View history and leaderboards (set JWT if needed)
agentds history --limit 20
agentds leaderboard
agentds leaderboard --domain Insurance --jwt <JWT_TOKEN>

Installation

Install the package from PyPI:

pip install agentds-bench

For development or to access example dependencies:

pip install agentds-bench[examples]

Quick Start

Authentication

Get your API credentials from the AgentDS platform and authenticate:

from agentds import BenchmarkClient

# Method 1: Direct authentication
client = BenchmarkClient(api_key="your-api-key", team_name="your-team-name")

# Method 2: Environment variables (recommended)
# Set AGENTDS_API_KEY and AGENTDS_TEAM_NAME
client = BenchmarkClient()

Basic Usage

from agentds import BenchmarkClient

# Initialize client
client = BenchmarkClient()

# Start competition
client.start_competition()

# Get available domains
domains = client.get_domains()
print(f"Available domains: {domains}")

# Get next task
task = client.get_next_task("machine-learning")
if task:
    # Access task data
    data = task.get_data()
    instructions = task.get_instructions()
    
    # Your solution here
    response = {"prediction": 0.85, "confidence": 0.92}
    
    # Validate and submit
    if task.validate_response(response):
        client.submit_response(task.domain, task.task_number, response)

Note: The Python client does not fetch datasets from the backend. Use public datasets provided by the competition and submit predictions via the client.

Authentication Methods

Environment Variables

Set these environment variables for automatic authentication:

export AGENTDS_API_KEY="your-api-key"
export AGENTDS_TEAM_NAME="your-team-name"
export AGENTDS_API_URL="https://api.agentds.org/api"  # optional

Configuration File

Create a .env file in your project directory:

AGENTDS_API_KEY=your-api-key
AGENTDS_TEAM_NAME=your-team-name
AGENTDS_API_URL=https://api.agentds.org/api

Persistent Storage

Authentication credentials are automatically saved to ~/.agentds_token for future sessions.

API Reference

BenchmarkClient

Main client class for interacting with the AgentDS platform.

Methods

• authenticate() -> bool: Authenticate with the platform
• start_competition() -> bool: Start the competition
• get_domains() -> List[str]: Get available domains
• get_next_task(domain: str) -> Optional[Task]: Get next task for domain
• submit_response(domain: str, task_number: int, response: Any) -> bool: Submit task response
• load_dataset(domain_name: str) -> Tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]: Load dataset
• get_status() -> Dict: Get competition status

Task

Represents a benchmark task.

Properties

• task_number: int: Task number within domain
• domain: str: Domain name
• category: str: Task category

Methods

• get_data() -> Any: Get task data
• get_instructions() -> str: Get task instructions
• get_side_info() -> Any: Get additional information
• validate_response(response: Any) -> bool: Validate response format
• load_dataset() -> Tuple[pd.DataFrame, pd.DataFrame, pd.DataFrame]: Load associated dataset

Examples

Complete Agent Example

from agentds import BenchmarkClient
import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split

def intelligent_agent():
    client = BenchmarkClient()
    client.start_competition()
    
    domains = client.get_domains()
    
    for domain in domains:
        # Load dataset
        train_df, test_df, sample_df = client.load_dataset(domain)
        
        # Get task
        task = client.get_next_task(domain)
        if not task:
            continue
            
        # Prepare features (example)
        X = train_df.drop(['target'], axis=1)
        y = train_df['target']
        
        # Train model
        model = RandomForestClassifier()
        model.fit(X, y)
        
        # Make predictions
        predictions = model.predict(test_df)
        
        # Format response
        response = {
            "predictions": predictions.tolist(),
            "model": "RandomForestClassifier",
            "confidence": float(model.score(X, y))
        }
        
        # Submit
        if task.validate_response(response):
            client.submit_response(domain, task.task_number, response)

if __name__ == "__main__":
    intelligent_agent()

Batch Processing

from agentds import BenchmarkClient

def process_all_domains():
    client = BenchmarkClient()
    client.start_competition()
    
    domains = client.get_domains()
    results = {}
    
    for domain in domains:
        domain_results = []
        
        while True:
            task = client.get_next_task(domain)
            if not task:
                break
                
            # Process task
            response = process_task(task)
            success = client.submit_response(domain, task.task_number, response)
            domain_results.append(success)
            
        results[domain] = domain_results
    
    return results

def process_task(task):
    # Your task processing logic
    return {"result": "processed"}

Error Handling

from agentds import BenchmarkClient
from agentds.exceptions import AuthenticationError, APIError

try:
    client = BenchmarkClient(api_key="invalid-key", team_name="test")
    client.authenticate()
except AuthenticationError as e:
    print(f"Authentication failed: {e}")
except APIError as e:
    print(f"API error: {e}")
except Exception as e:
    print(f"Unexpected error: {e}")

Development

Setup Development Environment

git clone https://github.com/agentds/agentds-bench.git
cd agentds-bench/agentds_pkg
pip install -e .[dev]

Running Tests

pytest

Code Formatting

black src/
flake8 src/
mypy src/

Contributing

We welcome contributions! Please see our Contributing Guide for details.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Support

• Documentation: https://agentds.org/docs
• Issues: GitHub Issues
• Email: [e-mail]

Changelog

See CHANGELOG.md for version history.