persona-bench 0.1.2

Pluristic alignment evaluation benchmark for LLMs

PERSONA Bench

Reproducible Testbed for Evaluating and Improving Language Model Alignment with Diverse User Values

SynthLabs.ai/research/persona

📄 Paper | 🗃️ Research Visualizations | 🤗 Hugging Face |

🌐 SynthLabs Research | 👥 Join the Team | 🤝 Let's Collaborate

PERSONA Bench is an extension of the PERSONA framework introduced in Castricato et al. 2024. It provides a reproducible testbed for evaluating and improving the alignment of language models with diverse user values.

Introduction

PERSONA established a strong correlation between human judges and language models in persona-based personalization tasks. Building on this foundation, we've developed a suite of robust evaluations to test a model's ability to perform personalization-related tasks. This repository provides practitioners with tools to assess and improve the pluralistic alignment of their language models.

Our evaluation suite uses inspect-ai to perform various assessments on persona-based tasks, offering insights into model performance across different demographic intersections, feature importance, and personalization capabilities.

Key Features

• 🎭 Main Evaluation: Assess personalized response generation
• 🧩 Leave One Out Analysis: Measure attribute impact on performance
• 🌐 Intersectionality: Evaluate model performance across different demographic intersections
• 🎯 Pass@K: Determine attempts needed for successful personalization
• 🔍 Comparison: Grounded personalization evaluation (API-exclusive)

Quick Start

1. Install Poetry if you haven't already:

   curl -sSL https://install.python-poetry.org | python3 -
   

2. Install the package:

   poetry add persona-bench
   

3. Use in your Python script:

   from dotenv import load_dotenv
   from persona_bench import evaluate_model
   
   # optional, you can also pass the environment variables directly to evaluate_model
   load_dotenv()
   
   eval = evaluate_model("gpt-3.5-turbo", evaluation_type="main")
   print(eval.results.model_dump())
   

PERSONA API

PERSONA Bench now offers an API for easy integration and evaluation of your models. The API provides access to all evaluation types available in PERSONA Bench, including a novel evaluation type called "comparison" for grounded personalization evaluation.

Quick Start with API

1. Install the package:

   pip install persona-bench
   

2. Set up your API key:

   • Sign up at https://www.synthlabs.ai/research/persona to get your API key and claim your free trial credits.
   • Set the API key as an environment variable:

     export SYNTH_API_KEY=your_api_key_here
     

3. Use in your Python script:

   from persona_bench.api import PERSONAClient
   from persona_bench.api.prompt_constructor import ChainOfThoughtPromptConstructor
   
   # Create a PERSONAClient object
   client = PERSONAClient(
       model_str="your_model_name",
       evaluation_type="comparison", # Run a grounded evaluation, API exclusive!
       N=50,
       prompt_constructor=ChainOfThoughtPromptConstructor(),
       # If not set as an environment variable, pass the API key here:
       # api_key="your_api_key_here"
   )
   
   # Iterate through questions and log answers
   for idx, q in enumerate(client):
       answer = your_model_function(q["system"], q["user"])
       client.log_answer(idx, answer)
   
   # Evaluate the results
   results = client.evaluate(drop_answer_none=True)
   print(results)
   

Key Features

• 🎭 Multiple Evaluation Types: Support for grounded, main, LOO, intersectionality, and pass@k evaluations
• 🔧 Customizable Prompt Construction: Use default or custom prompt constructors
• 📊 Easy Data Handling: Iterate through questions and log answers seamlessly
• 📈 Evaluation: Evaluate model performance with a single method call

Detailed Usage

Initialization

Create a PERSONAClient object with the following parameters:

• model_str: The identifier for this evaluation task
• evaluation_type: Type of evaluation ("main", "loo", "intersectionality", "pass_at_k", "comparison")
• N: Number of samples for evaluation
• prompt_constructor: Custom prompt constructor (optional)
• intersection: List of intersection attributes (required for intersectionality evaluation)
• loo_attributes: Leave-one-out attributes (required for LOO evaluation)
• seed: Random seed for reproducibility (optional)
• url: API endpoint URL (optional, default is "https://synth-api-development.eastus.azurecontainer.io/api/v1/personas/v1/")
• api_key: Your SYNTH API key (optional if set as an environment variable)

Iterating Through Questions

Use the client as an iterable to access questions:

for idx, question in enumerate(client):
    system_prompt = question["system"]
    user_prompt = question["user"]
    answer = your_model_function(system_prompt, user_prompt)
    client.log_answer(idx, answer)

Evaluation

Evaluate the logged answers:

results = client.evaluate(drop_answer_none=True, save_scores=False)

Advanced Usage

Custom Prompt Constructors

Create a custom prompt constructor by inheriting from BasePromptConstructor:

from persona_bench.api.prompt_constructor import BasePromptConstructor

class MyCustomPromptConstructor(BasePromptConstructor):
    def construct_prompt(self, persona, question):
        # Implement your custom prompt construction logic
        pass

client = PERSONAClient(
    # ... other parameters ...
    prompt_constructor=MyCustomPromptConstructor(),
)

Accessing Raw Data

Access the underlying data using indexing:

question = client[0]  # Get the first question

answers = [generate_answer(q) for q in client]
client.set_answers(answers)

Evaluation Types

Comparison Evaluation (API-exclusive)

The comparison evaluation is our most advanced and grounded assessment, exclusively available through the PERSONA API. It provides a robust measure of a model's personalization capabilities using known gold truth answers.

Click to expand details

• Uses carefully curated persona pairs with known distinctions
• Presents models with questions that have objectively different answers for each persona
• Evaluates the model's ability to generate persona-appropriate responses
• Compares model outputs against gold truth answers for precise accuracy measurement
• Offers the most reliable and interpretable results among all evaluation types

Example usage:

from persona_bench.api import PERSONAClient
client = PERSONAClient(model_str="your_identifier_name", evaluation_type="comparison", N=50)

Development Setup

1. Clone the repository:

   git clone https://github.com/SynthLabsAI/PERSONA-bench.git
   cd PERSONA-bench
   

2. Install dependencies:

   poetry install
   

3. Install pre-commit hooks:

   poetry run pre-commit install
   

4. Set up HuggingFace authentication:

   huggingface-cli login
   

5. Set up environment variables:

   cp .env.example .env
   vim .env
   

Detailed Evaluations

Main Evaluation

The main evaluation script assesses a model's ability to generate personalized responses based on given personas from our custom filtered PRISM dataset.

Click to expand details

1. Load PRISM dataset
2. Generate utterances using target model with random personas
3. Evaluate using GPT-4 as a critic model via a debate approach
4. Analyze personalization effectiveness

Leave One Out Analysis

This evaluation measures the impact of individual attributes on personalization performance.

Click to expand details

• Uses sub-personas separated by LOO attributes
• Tests on multiple personas and PRISM questions
• Analyzes feature importance

Available attributes include age, sex, race, education, employment status, and many more. See the leave one out example json for formatting.

The available attributes are

[
  "age",
  "sex",
  "race",
  "ancestry",
  "household language",
  "education",
  "employment status",
  "class of worker",
  "industry category",
  "occupation category",
  "detailed job description",
  "income",
  "marital status",
  "household type",
  "family presence and age",
  "place of birth",
  "citizenship",
  "veteran status",
  "disability",
  "health insurance",
  "big five scores",
  "defining quirks",
  "mannerisms",
  "personal time",
  "lifestyle",
  "ideology",
  "political views",
  "religion",
  "cognitive difficulty",
  "ability to speak english",
  "vision difficulty",
  "fertility",
  "hearing difficulty"
]

Example usage is:

from dotenv import load_dotenv
from persona_bench import evaluate_model

# optional, you can also pass the environment variables directly to evaluate_model
# make sure that your .env file specifies where the loo_json is!
load_dotenv()

eval = evaluate_model("gpt-3.5-turbo", evaluation_type="loo")
print(eval.results.model_dump())

Intersectionality

Evaluate model performance across different demographic intersections.

Click to expand details

• Define intersections using JSON configuration
• Measure personalization across disjoint populations
• Analyze model performance for specific demographic combinations

See the intersectionality example json.

This configuration defines two intersections:

Males aged 18-34 Females aged 18-34

You can use any of the attributes available in the LOO evaluation to create intersections. For attributes with non-enumerable values (e.g., textual background information), you may need to modify the intersection script to use language model embeddings for computing subpopulations.

Pass@K

Determines how many attempts are required to successfully personalize for a given persona.

Click to expand details

• Reruns main evaluation K times
• Counts attempts needed for successful personalization
• Provides insights into model consistency and reliability

WARNING! Pass@K is very credit intensive and may require multiple hours to complete a large run.

Running with InspectAI

Configure your .env file before running the scripts. You can set the generate mode to one of the following:

• baseline: Generate an answer directly, not given the persona
• output_only: Generate answer given the persona, without chain of thought
• chain_of_thought: Generate chain of thought before answering, given the persona
• demographic_summary: Generate a summary of the persona before answering

# Activate the poetry environment
poetry shell

# Main Evaluation
inspect eval src/persona_bench/main_evaluation.py --model {model}

# Leave One Out Analysis
inspect eval src/persona_bench/main_loo.py --model {model}

# Intersectionality Evaluation
inspect eval src/persona_bench/main_intersectionality.py --model {model}

# Pass@K Evaluation
inspect eval src/persona_bench/main_pass_at_k.py --model {model}

Using Inspect AI allows you to utilize their visualization tooling, which is found in their documentation here.

Visualization

We provide scripts for visualizing evaluation results:

• visualization_loo.py: Leave One Out analysis
• visualization_intersection.py: Intersectionality evaluation
• visualization_pass_at_k.py: Pass@K evaluation

These scripts use the most recent log file by default. Use the --log parameter to specify a different log file.

Dependencies

Key dependencies include:

• inspect-ai
• datasets
• pandas
• openai
• instructor
• seaborn

For development:

• tiktoken
• transformers

See pyproject.toml for a complete list of dependencies.

Citation

If you use PERSONA in your research, please cite our paper:

@misc{castricato2024personareproducibletestbedpluralistic,
      title={PERSONA: A Reproducible Testbed for Pluralistic Alignment},
      author={Louis Castricato and Nathan Lile and Rafael Rafailov and Jan-Philipp Fränken and Chelsea Finn},
      year={2024},
      eprint={2407.17387},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2407.17387},
}

Community & Support

Join our Discord community for discussions, support, and updates or reach out to us at https://www.synthlabs.ai/contact.

Acknowledgements

This research is supported by SynthLabs. We thank our collaborators and the open-source community for their valuable contributions.

---

Copyright © 2024, SynthLabs. Released under the Apache License.