openai/simple-evals

Features

• Language Model Benchmark Suites - Runs a predefined set of standardized tests to measure language model accuracy across tasks like reasoning, math, and coding.
• Unified Provider Interfaces - Generates responses from OpenAI and Anthropic APIs through a unified interface for standardized evaluation.
• Criteria-Based Scoring Engines - Sums points for criteria met, subtracts penalties, then divides by total possible points to compute a fractional score.
• Ground-Truth Scoring - Measures accuracy by comparing model outputs against curated answer keys using exact or fuzzy matching.
• Generative Model Sampling - Provides a unified interface for sampling responses from OpenAI and Claude language models.
• Language Model Response Generators - Sends conversation histories to language models and returns generated text with query metadata.
• Model Benchmarks - Runs a suite of standardized benchmarks to measure language model accuracy on reasoning, math, and coding.
• Multi-Provider Sampling Interfaces - Samples from OpenAI and Anthropic APIs through a unified interface for standardized evaluation.
• Model Benchmarking Tools - Runs standardized tests across multiple language model providers to compare performance metrics.
• Model Evaluation Frameworks - Provides a unified framework for running model inference and validation across standard language model benchmarks.
• Model Evaluation and Benchmarking - Measures language model performance on standardized benchmarks including math, coding, and factuality tasks.
• Standardized Benchmarks - Runs a suite of standardized benchmarks against a language model and reports accuracy scores for each test.
• Rubric-Based Graders - Implements a rubric-driven grading system that automatically scores model responses against expert-defined criteria.
• Rubric-Based Evaluators - Grades model completions against predefined rubric criteria, returning a boolean for each criterion met.
• Code Generation Benchmarks - Assesses model ability to produce correct code from natural language descriptions using established coding benchmarks.
• Code Generation Evaluators - Measures how often language models produce functionally correct code by running completions against test suites.
• Code Correctness Testings - Measures how often language models produce functionally correct code by running generated completions against test suites.
• Code Correctness Verifications - Verifies code correctness by running generated completions against predefined test suites with parallel execution.
• Domain Knowledge Evaluations - Measures model performance on specialized benchmarks including health, browsing comprehension, and graduate-level QA.
• Medical Knowledge Assessments - Evaluates language model accuracy on medical domain questions using specialized datasets like HealthBench.
• Evaluation Report Aggregators - Outputs per-evaluation HTML reports and JSON metric files for each combination of model sampler and language variant.
• Verbosity Penalties - Reduces scores proportionally when responses exceed a configurable character threshold with per-unit penalties.
• CLI Evaluation Runners - Launches benchmark evaluations for a specified model through a simple command-line invocation.
• Multilingual Accuracy Evaluations - Measures language model consistency and reasoning accuracy across multiple languages using translated benchmark datasets.
• Parallel Evaluators - Executes multiple code correctness checks concurrently using a thread pool to speed up batch evaluation.
• Factuality Benchmarking Frameworks - Tests model precision on fact-based question answering tasks against curated knowledge benchmarks.
• Benchmark Translations - Tests model consistency across languages by running translated versions of standard benchmarks like MMLU.
• Performance Metrics - Calculates quantified metrics such as precision, recall, F1 scores, and pass-at-k from evaluation results.
• Pass-at-K Calculators - Calculates the probability that at least one of K generated code samples passes all tests for robust accuracy measurement.
• Pass-at-K Statistical Scorings - Calculates the probability that at least one of K generated samples passes all tests for robust accuracy measurement.
• Chain-of-Thought Evaluations - Tests models using simple instructions without few-shot examples to better reflect realistic usage.
• Medical Knowledge Assessors - Ships a dedicated medical knowledge evaluator using the HealthBench dataset for accuracy scoring.
• Language Model Math Evaluations - Measures language model accuracy on mathematical reasoning tasks by prompting step-by-step solving and comparing answers against ground truth.
• Mathematical Reasoning Evaluations - Measures language model accuracy on mathematical problem-solving benchmarks using standardized test sets.
• Healthcare Knowledge Benchmarks - Runs models against the HealthBench dataset and reports accuracy scores for medical knowledge assessment.
• Evaluation Frameworks - Evaluation tools provided by OpenAI.
• Web and Environment Benchmarks - Challenging benchmark for browsing agents.

Istoric stele