LLM Steerability Evaluation

This repository contains an implementation of a "trade-off steerable benchmark" - a framework for evaluating how well AI systems can adapt to reflect different user perspectives and personalities. The benchmark includes:

1. A dataset of 6,000 statements across 100 diverse personas seeded from 5 personality frameworks (MBTI, Enneagram, Big Five, Zodiac, and Tarot).
2. An evaluation pipeline that measures how well LLMs can be steered to match different personas.
3. Tools for analyzing and visualizing system performance.

This work builds on recent research in pluralistic alignment [1] - the idea that AI systems should be able to reflect diverse human values rather than being aligned to a single set of preferences. Our implementation is inspired by Sorensen et al.'s proposal for "trade-off steerable benchmarks" and draws on techniques from Anthropic's work on model-written evaluations [2] for dataset generation and validation.

Key Findings

Our initial experiments with few-shot steerable systems showed:

• Even simple few-shot steering can produce meaningful persona adaptation, with most models achieving >80% steerability scores
• Claude 3.5 Sonnet achieved the strongest performance (94.6% steerability), followed by GPT-4o Mini (89.9%) and Gemini 1.5 Flash (80.2%)
• Models showed clear ability to maintain distinct behavior patterns while adapting to different personas
• Natural clustering emerged between similar personas across frameworks

Getting Started

Note: This project is still under active development. Some of the code isn't beautiful, and there is old code lying around. We're working on cleaning it up - in the meantime, proceed with caution and let us know if you run into any issues.

We strongly recommend that you create a Python virtual environment to manage dependencies. After you've done this, install the dependencies:

pip install -r requirements.txt

Creating the Dataset

The dataset generation pipeline creates personality-aligned statements using LLMs with filtering for quality and diversity:

Copy local.env.template to local.env and set your API keys.

cp local.env.template local.env

Run the dataset generation script. From the root directory, run:

python -m scripts.create_dataset

Running the Evaluation

Copy config_template.json to my_eval.json and set your evaluation parameters.

cp configs/config_template.json configs/my_eval.json

{
    "experiment_name": "2024-12-11-claude-40-1",
    "resume": true,
    "run_async": true,
    "restore_async": true,
    "max_concurrent_tests": 10,
    "max_concurrent_steering_tasks": 8,
    "personas_path": "dataset/personas_all_frameworks_2024-12-04.csv",
    "observations_path": "dataset/statements_all_frameworks_30_2024-12-04.csv",
    "max_personas": 40,
    "random_state": 42,
    "n_steer_observations_per_persona": 4,
    "inference_batch_size": 10,
    "batched_inference": false,
    "steerable_system_type": "FewShotSteerable",
    "steerable_system_config": {
        "llm_provider": "anthropic",
        "model": "claude-3-5-sonnet-latest",
        "verbose": true
    },
    "verbose": true,
    "output_base_dir": "output/experiments"
} 

Run the evaluation. From the root directory, run:

python -m scripts.run_statements_eval configs/my_eval.json

Citing this work

@misc{steerable-benchmark-2024,
title={LLM Steerability Evaluation},
author={Plastic Labs},
year={2024},
howpublished={\url{https://github.com/plastic-labs/steerability-eval}}
}

References

1. T. Sorensen, J. Moore, J. Fisher, M. Gordon, N. Mireshghallah, C. M. Rytting, A. Ye, L. Jiang, X. Lu, N. Dziri, T. Althoff, and Y. Choi, "A Roadmap to Pluralistic Alignment," arXiv preprint arXiv:2402.05070, 2024.
2. E. Perez, S. Ringer, K. Lukošiūtė, K. Nguyen, et al., "Discovering Language Model Behaviors with Model-Written Evaluations," arXiv preprint arXiv:2212.09251, 2022.