Toolkit Eval Harness

Enterprise-grade evaluation framework for AI/ML models with versioned test suites, deterministic scoring, and CI/CD integration for regression testing.

Overview

The Toolkit Eval Harness is a lightweight, powerful evaluation tool designed for AI/ML engineers to version "golden" task suites, score predictions with deterministic metrics, and produce CI-friendly JSON reports with regression comparisons. It complements Neural Forge model lifecycle gates and provides a solid foundation for model evaluation in production environments.

Key Features

Test Suite Management

• Versioned Test Suites: Immutable, version-controlled evaluation datasets
• Deterministic Scoring: Consistent, reproducible evaluation metrics
• Suite Packs: Compressed, hash-verified test suite packages
• Digital Signatures: Optional cryptographic signing for integrity verification

Comprehensive Evaluation

• Multiple Scoring Methods: Exact match, JSON schema validation, custom metrics
• Flexible Test Cases: Support for various input/output formats
• Tagging System: Organize tests by category, difficulty, or use case
• Batch Processing: Efficient evaluation of large test suites

Enterprise Integration

• CI/CD Friendly: JSON reports with exit codes for pipeline integration
• Regression Detection: Automated comparison against baseline evaluations
• Audit Trails: Complete evaluation history and provenance tracking
• Performance Metrics: Latency, throughput, and resource usage tracking

Security and Compliance

• Package Signing: Ed25519 cryptographic signatures for integrity
• Hash Verification: SHA-256 checksums for all packages
• Access Control: Role-based permissions for suite management
• Audit Logging: Complete evaluation audit trails

Quick Start

Installation

# Install from source
git clone https://github.com/AKIVA-AI/toolkit-eval-harness.git
cd toolkit-eval-harness
pip install -e ".[dev]"

# Install with signing support
pip install -e ".[signing]"

# Install in production
pip install toolkit-eval-harness

Basic Usage

# 1. Create a test suite pack
toolkit-eval pack create --suite-dir examples/suite --out packs/suite.zip

# 2. Verify pack integrity
toolkit-eval pack verify --suite packs/suite.zip

# 3. Run evaluation
toolkit-eval run --suite packs/suite.zip --predictions examples/preds.jsonl --out report.json

# 4. Compare with baseline (CI gating)
toolkit-eval compare --baseline baseline.json --candidate report.json

CLI Commands

• pack create - Create a suite pack from a directory
• pack verify - Verify pack integrity (hashes)
• pack sign - Sign suite packs (optional)
• pack verify-signature - Verify pack signatures
• run - Run evaluation against predictions
• compare - Compare candidate report to baseline (CI gating)
• keygen - Generate signing keys

Exit Codes

• 0 - Success
• 1 - General error
• 4 - Regression detected (for compare command)

License

MIT License - see LICENSE file for details.