polaudit

An audit framework for LLM alignment with human subgroups, based on polarization. Whose perspective does an LLM's label reflect? When people disagree about a text, a language model's single label sides with someone. polaudit finds the texts where annotators are polarized, finds which annotator characteristic (gender, politics, age, ...) explains the split, and measures which side each model takes.

The audit in four steps

1. Detect polarization. Each text gets a DFU score (distance from unimodality). A text with DFU above 0 is polarized.
2. Attribute it. A polarized text is attributed to a characteristic when every group of that characteristic is internally unimodal (a-posteriori unimodality, AU). Groups in fewer than 10 attributed texts are removed.
3. Ask the models. Your prompt, your models, optionally with a persona.
4. Measure alignment. Cohen's kappa between the model's labels and each group's position, with bootstrap intervals and a chance check.

Install

pip install polaudit                                  # the AU analysis
pip install "polaudit[plots,openai,gemini]"           # figures and the model providers you use

Try it in a minute (no data, no keys)

from polaudit import run_dataset
from polaudit.toy import toy_spec, toy_ratings

result = run_dataset(toy_spec(), raw=toy_ratings())
print(result.describe())
result.per_dimension        # attributed texts per characteristic

Your own data, prompt and model

from polaudit import DatasetSpec, ModelHub, annotate, load_model_inputs, run_dataset, save_au_tables

spec = DatasetSpec(
    name="mydata",
    load=lambda root: my_dataframe,        # columns: item_id, text, rating, plus one column per characteristic
    dimensions={"gender": ("gender", {"m": "man", "f": "woman"}),
                "politics": ("politics", {"l": "left", "r": "right"})},
)
result = run_dataset(spec, raw=my_dataframe, min_au_texts=10)     # min_au_texts=0 keeps all AU texts
save_au_tables({"mydata": result.au}, "results")

def my_prompt(texts, item):
    return f"Rate how toxic each {item} is from 1 to 5.\nFormat: score=<n>, label=<toxic/civil>\n\n{texts}"

annotate(ModelHub.from_env()["chatgpt"], load_model_inputs("results/aposteriori_for_models.csv"), "results", prompt=my_prompt, run="baseline")

Full documentation

This page is the short version. The full documentation lives on GitHub, with nine tutorials (a short document and a runnable notebook each), every parameter, and the code that reproduces the paper:

• Full README on GitHub: overview, repository layout, all tutorials, reproducing the paper
• Tutorials: documents and notebooks
• All parameters
• Reference: API keys, models, prompts, outputs, troubleshooting

The paper

This package is the code of Polarization Auditing Framework for LLMs: a Toxicity Detection Study (authors anonymous for review) and follows the current version of the paper.