evalbench 0.1.4

A collection of evaluation metrics for LLM tasks

EvalBench 🧪📊

Lightweight, extensible, open-source evaluation framework for LLM applications covering approximately 18 metrics and allows adding of user-defined custom metrics

---

🔍 About EvalBench

EvalBench is a plug-and-play Python package for evaluating outputs of large language models (LLMs) across a variety of metrics.

Supports:

• Predefined metrics (faithfulness, coherence, BLEU, hallucination, etc.)
• Easy registration of custom user-defined metrics
• Evaluation across different modalities/ use-cases
• Flexible input/output options (print/ save) with support for list-based batching

Modules and Metric Categories:

Module	Metrics	Required Arguments	Argument Types
response_quality	conciseness_score, coherence_score, factuality_score	response	List[str]
reference_based	bleu_score, rouge_score, meteor_score, semantic_similarity_score, bert_score	reference, generated	List[str], List[str]
contextual_generation	faithfulness_score, hallucination_score, groundedness_score	context, generated	List[List[str]], List[str]
retrieval	recall_at_k, precision_at_k, ndcg_at_k, mrr_score	relevant_docs, retrieved_docs, k	List[List[str]], List[List[str]], int
query_alignment	context_relevance_score	query, context	List[str], List[str]
response_alignment	answer_relevance_score, helpfulness_score	query, response	List[str], List[str]
user defined module	User-registered custom metrics	Varies (user-defined)	Varies (user-defined)

EvalBench is especially useful when you're:

• Building small-scale LLM pipelines
• Comparing different prompts or model outputs
• Rapidly iterating on metrics

---

🚀 Usage

Installation

pip install evalbench

Initialize Configuration

import evalbench as eb

# Create and apply evaluation configuration
config = eb.EvalConfig(groq_api_key="", output_mode="print")
eb.set_config(config)

Usage Examples

1. Evaluate a single predefined metric

response = ["A binary search algorithm reduces the time complexity to O(log n)."]
context = [["Binary search works on sorted arrays and is faster than linear search."]]

eb.faithfulness_score(context=context, generated=response)

# metrics can also be printed as a list
result = eb.faithfulness_score(context=context, generated=response)
print(result) # [0.44]

2. Evaluate all metrics in a predefined module

eb.evaluate_module(
    module=["contextual_generation"],
    context=context,
    generated=response,
)

3. Register custom metrics

Create a custom metric in a file (eg. custom.py):

from evalbench import register_metric, handle_output

@register_metric(name="len_metric", module="custom", required_args=["response", "reference"])
@handle_output()
def my_custom_metric(response, reference):
    return [len(response) / len(reference)]

4. Load custom metric file and evaluate

eb.load_custom_metrics("custom.py")

response = ["A binary search algorithm reduces the time complexity to O(log n).", "The Eiffel Tower is located in Berlin and was built in the 1800s."]
reference = ["Binary search works on sorted arrays and is faster than linear search.", "In Python, a generator yields items one at a time using the 'yield' keyword."]

# Evaluate custom metric directly
eb.my_custom_metric(response, reference)

# Evaluate all metrics in the 'custom' module
eb.evaluate_module(
    module=["custom"],
    response=response,
    reference=reference,
)

---

💡 Use Cases

EvalBench is built for fast feedback loops while developing:

• LLM Applications: Chatbots, assistants, summarization tools
• Prompt Engineering: Compare prompt variations using faithfulness, conciseness, or BLEU
• Model Evaluation: Benchmark outputs from different model runs
• Custom Evaluation Design: Rapidly prototype domain-specific metrics

---

🚧 Coming Soon

• Basic visualizations: Histograms and score distributions to quickly interpret results
• Batch mode via JSONL/Pandas: Evaluate and export results from structured files