evalx 0.1.0

Next-generation evaluation framework for LLM applications with research-grade validation and production-ready performance

EvalX: Next-Generation LLM Evaluation Framework

EvalX is a comprehensive evaluation framework for Large Language Model applications that combines traditional metrics, LLM-as-judge evaluations, and intelligent agentic orchestration with research-grade validation.

🚀 Key Features

• 🤖 Agentic Orchestration: Natural language instructions → automatic evaluation planning
• 📊 Comprehensive Metrics: Traditional + LLM-as-judge + hybrid approaches
• 🔬 Research-Grade Validation: Statistical analysis, confidence intervals, meta-evaluation
• 🎨 Multimodal Support: Vision-language, code, audio evaluation
• ⚡ Production Ready: Async processing, caching, CLI interface
• 🎯 Adaptive Selection: AI-powered optimal metric selection

🏗️ Unique Innovations

Meta-Evaluation System

EvalX includes the industry's first meta-evaluation system that assesses the quality of evaluation metrics themselves:

• Reliability assessment through test-retest analysis
• Validity measurement against ground truth
• Bias detection across demographic groups
• Interpretability scoring

Adaptive Metric Selection

Automatically selects optimal metrics based on:

• Task type and domain
• Quality requirements (research vs. production)
• Computational constraints
• Fairness requirements

📦 Installation

pip install evalx

For development:

pip install evalx[dev]

For research features:

pip install evalx[research]

For production deployment:

pip install evalx[production]

🎯 Quick Start

Natural Language Evaluation

import evalx

# Create evaluation suite from natural language instruction
suite = evalx.EvaluationSuite.from_instruction(
    "Evaluate my chatbot responses for helpfulness and accuracy"
)

# Your data
data = [
    {
        "input": "What's the capital of France?",
        "output": "The capital of France is Paris.",
        "reference": "Paris is the capital city of France."
    }
]

# Run evaluation
results = await suite.evaluate_async(data)
print(results.summary())

Fine-Grained Control

from evalx import MetricSuite

# Create custom metric combination
suite = MetricSuite()
suite.add_traditional_metric("bleu_score")
suite.add_traditional_metric("semantic_similarity", threshold=0.8)
suite.add_llm_judge("accuracy", model="gpt-4")

results = suite.evaluate(data)

Research-Grade Analysis

from evalx import ResearchSuite

# Comprehensive statistical analysis
suite = ResearchSuite(
    metrics=["accuracy", "helpfulness", "bleu"],
    confidence_level=0.95,
    bootstrap_samples=1000
)

results = await suite.evaluate_research_grade(data)
print(f"Mean ± Std: {results.mean:.3f} ± {results.std:.3f}")
print(f"95% CI: [{results.confidence_interval[0]:.3f}, {results.confidence_interval[1]:.3f}]")

🎨 Multimodal Evaluation

from evalx.metrics.multimodal import MultimodalInput, ImageCaptionQualityMetric

# Image captioning evaluation
input_data = MultimodalInput(
    input_text="Describe this image",
    output_text="A beautiful sunset over the ocean",
    image="path/to/image.jpg"
)

metric = ImageCaptionQualityMetric()
result = metric.evaluate(input_data)

🔬 Meta-Evaluation

from evalx.meta_evaluation import MetaEvaluator

# Evaluate your metrics' quality
meta_evaluator = MetaEvaluator()
quality_report = meta_evaluator.evaluate_metric_quality(
    metric=my_metric,
    evaluation_data=test_data,
    ground_truth=human_ratings
)

print(f"Metric Quality: {quality_report.overall_quality:.3f}")
print(f"Reliability: {quality_report.reliability:.3f}")
print(f"Validity: {quality_report.validity:.3f}")
print(f"Bias Score: {quality_report.bias:.3f}")

🖥️ Command Line Interface

# Evaluate using natural language
evalx evaluate "Check my chatbot for helpfulness" --data data.json

# Research-grade evaluation
evalx research --data data.json --metrics accuracy helpfulness --confidence 0.95

# List available metrics
evalx metrics --list

📊 Supported Metrics

Traditional Metrics

• BLEU: N-gram overlap with smoothing
• ROUGE: Recall-oriented evaluation (ROUGE-1, ROUGE-2, ROUGE-L)
• METEOR: Semantic matching with synonyms and stemming
• BERTScore: Contextual embedding similarity
• Semantic Similarity: Sentence transformer-based
• Exact Match: String matching with normalization
• Levenshtein: Edit distance with word/character level

LLM-as-Judge Metrics

• Accuracy: Factual correctness assessment
• Helpfulness: Response utility evaluation
• Coherence: Logical consistency measurement
• Groundedness: Source attribution verification
• Relevance: Query-response alignment

Multimodal Metrics

• Image-Text Alignment: CLIP-based similarity
• Image Caption Quality: Comprehensive captioning assessment
• Code Correctness: Syntax, execution, security analysis
• Audio Quality: Signal processing metrics

🏆 Why EvalX?

Feature	EvalX	DeepEval	LangChain	Ragas
Meta-evaluation	✅ Unique	❌	❌	❌
Statistical rigor	✅ Best	Basic	Basic	Good
Multimodal support	✅ Comprehensive	Limited	Limited	Limited
Adaptive selection	✅ Unique	❌	❌	❌
Natural language interface	✅ Full	❌	❌	❌
Production ready	✅ Complete	Good	Basic	Good

📚 Documentation

• Architecture Overview
• Installation Guide
• Examples
• API Reference

🤝 Contributing

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

📄 License

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

🙏 Acknowledgments

• Built for the AI evaluation community
• Inspired by advances in LLM evaluation research
• Designed for both researchers and practitioners

📞 Support

• GitHub Issues
• Documentation
• Community Discussions

---

EvalX: Making AI evaluation comprehensive, reliable, and accessible.