AlpacaEval logo

AlpacaEval

Free

An Automatic Evaluator for Instruction-following Language Models using Nous benchmark suite.

FreeFree tier
Type
Open Source

About AlpacaEval

AlpacaEval is an LLM-based automatic evaluation system for instruction-following language models, designed to be fast, cheap, and reliable. It leverages the AlpacaFarm evaluation set, which consists of diverse but generally simple instructions, and uses GPT-4-based auto-annotators to compare model responses against reference responses (Davinci003 for AlpacaEval 1.0, GPT-4 Preview for AlpacaEval 2.0). The system calculates win rates and supports length-controlled (LC) win rates to mitigate length bias in annotations. AlpacaEval demonstrates high agreement with human annotations, and its leaderboard rankings correlate strongly with human-based evaluations. The tool is open-source, welcomes community contributions for new models, evaluators, and evaluation sets, and includes a Discord for support. Limitations include a focus on simple instructions, lack of safety evaluation, and residual length biases.

Key Features

LLM-based automatic evaluation using GPT-4 auto-annotators
Based on the AlpacaFarm evaluation set for diverse instruction-following tests
Calculates win rates compared to reference models (Davinci003 or GPT-4 Preview)
Length-controlled (LC) win rates to alleviate length bias
High agreement with human annotations and correlated leaderboard rankings
Open-source with community contribution support for models, evaluators, and eval sets
Leaderboard for transparent comparison of instruction-following models

Pros & Cons

Pros
  • Fast, cheap, and reliable automatic evaluation
  • High correlation with human annotations, providing trustworthy rankings
  • Open-source and actively maintained with community contributions
  • Length-controlled win rates mitigate length bias in evaluations
Cons
  • Does not evaluate model safety
  • Evaluation set consists mainly of simple instructions, limiting generalizability
  • Auto-annotator win rates still correlated with length despite LC mitigation
  • Not a comprehensive or gold-standard evaluation; should be used alongside other methods

Best For

Benchmarking and comparing instruction-following LLMsEvaluating model alignment with human preferencesResearch in model evaluation and automated annotationCommunity-driven model assessment and leaderboard participation

FAQ

What is AlpacaEval?
AlpacaEval is an LLM-based automatic evaluation system for instruction-following language models. It uses GPT-4 auto-annotators and the AlpacaFarm evaluation set to compute win rates and compare models on a leaderboard.
How does AlpacaEval work?
It compares model responses to reference responses (Davinci003 for version 1.0, GPT-4 Preview for version 2.0) using GPT-4-based auto-annotators. The system calculates win rates and optionally uses length-controlled win rates to reduce length bias.
What is the length-controlled (LC) win rate?
Length-controlled win rates are a modification that alleviates the bias of GPT-4 favoring longer outputs. However, they may favor models fine-tuned on GPT-4 outputs.
How can I add a new model to the leaderboard?
You need to run your model on the AlpacaEval evaluation set, auto-annotate the outputs, and submit a pull request with the model config and leaderboard results. Instructions are in the contributions section of the website.
Does AlpacaEval evaluate safety?
No, AlpacaEval does not evaluate the safety of models. It focuses on instruction-following capabilities.