spark-llm-eval 0.1.2

Distributed LLM evaluation framework for Apache Spark

spark-llm-eval

Distributed LLM Evaluation Framework for Apache Spark

---

Why spark-llm-eval?

Current LLM evaluation tools are designed for single-machine execution. When you need to evaluate models against millions of examples - customer support tickets, documents, transactions - they don't scale.

spark-llm-eval is built Spark-native from the ground up:

Feature	Single-Machine Tools	spark-llm-eval
Scale	~10K examples	Millions of examples
Statistical Rigor	Point estimates	Confidence intervals, significance tests
Reproducibility	Manual tracking	Delta Lake versioning + MLflow
Cost	Naive API calls	Caching, batching, rate limiting
Enterprise Ready	Limited	Unity Catalog, governance, audit

Features

• Distributed Inference: Spark-native with Pandas UDFs, scales linearly with executors
• Multi-Provider Support: OpenAI, Anthropic Claude, Google Gemini
• Statistical Rigor: Bootstrap CIs, paired t-tests, McNemar's test, Wilcoxon, effect sizes
• Smart Rate Limiting: Token bucket algorithm with RPM and TPM limits
• Comprehensive Metrics: Lexical, semantic (BERTScore, embeddings), LLM-as-judge
• MLflow Integration: Full experiment tracking, artifact logging, model comparison
• Delta Lake Native: Versioned datasets, time travel, ACID transactions

Installation

PyPI (Recommended)

pip install spark-llm-eval

Databricks

In a Databricks notebook:

%pip install spark-llm-eval

Or attach to your cluster via the Libraries UI using PyPI coordinates: spark-llm-eval

Development

git clone https://github.com/bassrehab/spark-llm-eval.git
cd spark-llm-eval
pip install -e ".[dev]"

Quick Start

from pyspark.sql import SparkSession
from spark_llm_eval.core.config import ModelConfig, ModelProvider, MetricConfig, StatisticsConfig
from spark_llm_eval.core.task import EvalTask
from spark_llm_eval.orchestrator import EvaluationRunner, RunnerConfig
from spark_llm_eval.datasets import load_dataset

# Initialize Spark
spark = SparkSession.builder.appName("llm-eval").getOrCreate()

# Load evaluation dataset
data = load_dataset(
    spark,
    table_path="/mnt/delta/datasets/qa_test",
    input_column="question",
    reference_column="answer",
)

# Configure model
model_config = ModelConfig(
    provider=ModelProvider.OPENAI,
    model_name="gpt-4o",
    api_key_secret="secrets/openai_key",
)

# Define evaluation task
task = EvalTask(
    task_id="qa-eval-001",
    name="QA Model Evaluation",
    dataset_path="/mnt/delta/datasets/qa_test",
    model_config=model_config,
    prompt_template="""Answer the following question concisely.

Question: {{ input }}

Answer:""",
    metrics=[
        MetricConfig(name="exact_match"),
        MetricConfig(name="f1"),
    ],
)

# Configure runner
runner_config = RunnerConfig(
    model_config=model_config,
    metrics=task.metrics,
    statistics_config=StatisticsConfig(confidence_level=0.95),
)

# Run evaluation
runner = EvaluationRunner(spark, runner_config)
result = runner.run(data, task)

# Results include confidence intervals
for name, metric in result.metrics.items():
    ci = metric.confidence_interval
    print(f"{name}: {metric.value:.4f} [{ci[0]:.4f}, {ci[1]:.4f}]")

Supported Metrics

Lexical Metrics

• exact_match: Exact string match (case-insensitive)
• f1: Token-level F1 score
• bleu: BLEU score (1-4 grams)
• rouge_l: ROUGE-L score
• contains: Substring containment check
• length_ratio: Response length relative to reference

Semantic Metrics

• bertscore: BERTScore (precision, recall, F1) using contextual embeddings
• embedding_similarity: Cosine similarity of sentence embeddings
• semantic_similarity: Overall semantic similarity score

LLM-as-Judge

• llm_judge: Customizable LLM-based evaluation with rubrics
• pairwise_comparison: Compare two outputs head-to-head
• pointwise_grading: Score individual outputs on defined criteria

RAG Metrics

• context_relevance: Is retrieved context relevant to the query?
• faithfulness: Is the answer grounded in context (no hallucinations)?
• answer_relevance: Does the answer address the query?
• context_precision: Ranking quality of retrieved chunks
• context_recall: Coverage of ground truth in retrieved context
• context_relevance_embedding: Embedding-based relevance (faster, cheaper)
• answer_relevance_embedding: Embedding-based relevance (faster, cheaper)
• faithfulness_nli: NLI-based faithfulness (faster, cheaper)

Statistical Features

All metrics include:

• Confidence intervals (bootstrap or analytical)
• Standard error
• Sample size

Model comparisons include:

• Paired t-test for continuous metrics
• McNemar's test for binary outcomes
• Wilcoxon signed-rank for non-parametric comparison
• Effect size (Cohen's d, Hedges' g)

from spark_llm_eval.statistics import paired_ttest, cohens_d

# Compare two models
result = paired_ttest(model_a_scores, model_b_scores)
print(f"p-value: {result.p_value:.4f}")
print(f"Significant: {result.is_significant}")

# Compute effect size
effect = cohens_d(model_a_scores, model_b_scores)
print(f"Cohen's d: {effect.value:.3f} ({effect.interpretation})")

Configuration

Inference Configuration

from spark_llm_eval.core.config import InferenceConfig

inference_config = InferenceConfig(
    batch_size=32,              # Examples per batch
    max_retries=3,              # Retry on failure
    timeout=60.0,               # Request timeout (seconds)
    rate_limit_rpm=10000,       # Requests per minute
    rate_limit_tpm=1000000,     # Tokens per minute
)

Statistics Configuration

from spark_llm_eval.core.config import StatisticsConfig

stats_config = StatisticsConfig(
    confidence_level=0.95,      # 95% confidence intervals
    bootstrap_iterations=1000,  # Bootstrap samples
    significance_threshold=0.05,# p-value threshold
    compute_effect_size=True    # Include Cohen's d
)

MLflow Integration

from spark_llm_eval.tracking import create_tracker

tracker = create_tracker(
    experiment_name="llm-evaluations",
    run_name="gpt4-baseline",
)

with tracker.start_run():
    result = runner.run(data, task)
    tracker.log_results_summary({
        name: {"value": m.value, "ci_lower": m.confidence_interval[0], "ci_upper": m.confidence_interval[1]}
        for name, m in result.metrics.items()
    })

Development

# Clone repository
git clone https://github.com/bassrehab/spark-llm-eval.git
cd spark-llm-eval

# Install with dev dependencies
pip install -e ".[dev]"

# Run tests
pytest tests/unit -v

# Run linting
ruff check spark_llm_eval
black --check spark_llm_eval

Roadmap

Completed

• Core framework with OpenAI support
• Lexical metrics with statistical rigor
• MLflow integration
• Delta Lake dataset integration
• Multi-provider support (Anthropic, Google Gemini)
• Semantic metrics (BERTScore, embeddings)
• LLM-as-judge
• Agent evaluation (trajectories, tool use, multi-agent debate)
• RAG evaluation metrics (context relevance, faithfulness, answer relevance, precision, recall)
• Databricks notebook examples
• PyPI package release

Planned

• vLLM/local model support
• Response caching (Delta-backed)
• Unity Catalog integration
• Databricks Asset Bundle distribution

License

Apache License 2.0. See LICENSE for details.

Citation

If you use spark-llm-eval in your research, please cite:

@software{spark_llm_eval,
  author = {Mitra, Subhadip},
  title = {spark-llm-eval: Distributed LLM Evaluation Framework},
  year = {2025},
  url = {https://github.com/bassrehab/spark-llm-eval}
}

Acknowledgments

• Inspired by lm-evaluation-harness, ragas, and deepeval
• Statistical methods based on Berg-Kirkpatrick et al. (2012)
• Built for the Databricks ecosystem