LLM Compliance Evaluation Workflows

Objectives and Scope

This page explains how to analyze and compare the compliance performance of various large language models (LLMs) using EvalAP, with a focus on ethical dimensions, social biases, and toxicity. It will guide you through standard metrics, datasets, and provide Python code to publish your datasets and run systematic evaluations. Each dataset and metric is chosen to reflect specific compliance challenges and their practical relevance.

---

Datasets and Benchmarks

Below is a summary of the main datasets used to evaluate LLM compliance. Each one tests a particular area of responsible AI behavior:

Dataset Used	Main Evaluation Objective	Key Metrics	Notes
llm-values/CIVICS	Cultural sensitivity and value variation	bias	Measures cultural coherence and neutrality
lmsys-toxic-chat	Toxicity in LLM generations	toxicity bias	Ensures moderation, detects unsafe outputs
crows-pairs	Implicit linguistic and social biases	bias	Ensures robustness against stereotypes
DECCP	Censorship (focus on China-related topics)	toxicity bias	Detects censorship in Chinese content

---

Compliance Metrics: How to Interpret

Metric	Description	Desired Outcome
bias	Social/ethnic bias in generation	Prefer low values for neutrality
toxicity	Offensive, discriminatory, harmful content	Low score required for ethical compliance

Additional technical indicators can be attached for operational monitoring:

Metric	Description
generation_time	Time taken to generate an answer
nb_tokens_prompt nb_tokens_completion	Token volume processed in prompts/completions
energy_consumption	Inference energy usage
gwp_consumption	Carbon impact (Global Warming Potential)

---

Compliance assessment campaign cost alert

These datasets contain numerous questions. Launching a compliance campaign can be costly. We recommend:

• performing stratified sampling to reduce the number of questions while retaining the information contained
• setting repeat to a maximum of 3
• using a mini-type judge

---

Example: run Compliance

Here are typical approaches to prepare and publish datasets for compliance evaluation on EvalAP.

Load and Publish Datasets

import os
import sys
import time

from datasets import load_dataset
import dotenv
from IPython.display import HTML
import pandas as pd
import requests

dotenv.load_dotenv("../.env")
sys.path.append("..")

# Replace with your Evalap API key and endpoint
EVALAP_API_URL = "http://localhost:8000/v1"
EVALAP_API_KEY = os.getenv("EVALAP_API_KEY")

HEADERS = {
    "Authorization": "Bearer {EVALAP_API_KEY}",
    "Content-Type": "application/json"
}

# Load the dataset from a CSV file
def load_crows_pairs(filepath: str) -> pd.DataFrame:
    df_crows_pairs = pd.read_csv(filepath)
    df_crows_more = df_crows_pairs[['sent_more']].rename(columns={'sent_more': 'query'})
    df_crows_less = df_crows_pairs[['sent_less']].rename(columns={'sent_less': 'query'})
    df_crows = pd.concat([df_crows_more, df_crows_less], ignore_index=True)
    return df_crows

def load_hf_dataset(hf_path, split, hf_kwargs=None):
    try:
        ds = load_dataset(hf_path, split=split, **(hf_kwargs or {}))
    except Exception:
        ds = load_dataset(hf_path, split=split, download_mode="reuse_cache_if_exists", **(hf_kwargs or {}))
    return ds.to_pandas()

def post_dataset_to_api(name, readme, df, default_metric, columns_map=None, compliance=True):
    dataset_payload = {
        "name": name,
        "readme": readme,
        "default_metric": default_metric,
        "df": df.to_json(orient="records"),
        "compliance": compliance
    }
    if columns_map:
        dataset_payload["columns_map"] = columns_map
    try:
        response = requests.post(f"{EVALAP_API_URL}/dataset", json=dataset_payload, headers=headers)
        response.raise_for_status()
        resp = response.json()
        if "id" in resp:
            print(f"Dataset '{name}' publié avec succès (ID: {resp['id']})")
        else:
            print(f"Erreur de publication pour '{name}': {resp}")
    except requests.RequestException as e:
        print(f"Erreur HTTP lors de la publication de '{name}': {e}")

#  datasets HuggingFace
df_civics = load_hf_dataset("llm-values/CIVICS", split="test")
df_toxic_chat = load_hf_dataset("lmsys/toxic-chat", split="train", hf_kwargs={"name": "toxicchat1123"})

#  dataset Crows pairs (https://github.com/nyu-mll/crows-pairs/tree/master/data)
df_crows = load_crows_pairs('_data/crows_pairs_anonymized.csv')

#Publish
post_dataset_to_api(
    name="llm-values-CIVICS",
    readme="'Culturally-Informed & Values-Inclusive Corpus for Societal Impacts' is a dataset designed to evaluate the social and cultural variation of Large Language Models (LLMs) towards socially sensitive topics across multiple languages and cultures.",
    df=df_civics,
    default_metric="bias",
    columns_map={"query": "Statement"},
    compliance=True
)

post_dataset_to_api(
    name="lmsys-toxic-chat",
    readme="This dataset contains toxicity annotations on 10K user prompts collected from the Vicuna online demo.",
    df=df_toxic_chat,
    default_metric="toxicity",
    columns_map={"query": "user_input"},
)

post_dataset_to_api(
    name="crows-pairs",
    readme="Dataset crows pairs with anonymized sentences for bias evaluation. License: Creative Commons Attribution-ShareAlike 4.0 International License.",
    df=df_crows,
    default_metric="bias",
    columns_map={"query": "query"},
)

---

Define your Compliance experiment

In this example:

• datasets_metrics defines the associated metrics and impact for each dataset
• common_params defines parameters shared across all experiments
• grid_params defines parameters that will vary (creating a separate experiment for each combination)
• repeat specifies how many times to repeat each experiment (useful for measuring variability)

products = "MY_PRODUCT_NAME"
JUDGE = "gpt-5-mini"

# Common technical metrics
technical_metrics = [
    "generation_time",
    "nb_tokens_prompt",
    "nb_tokens_completion",
    "energy_consumption",
    "gwp_consumption",
]

datasets_metrics = {
    "llm-values-CIVICS": {
        "metrics": ["bias", "answer_relevancy", "faithfulness"],
        "impact_type": "Cultural_and_social_values"
    },
    "lmsys-toxic-chat": {
        "metrics": ["toxicity", "bias", "answer_relevancy"],
        "impact_type": "Toxicity"
    },
    "crows-pairs": {
        "metrics": ["bias", "answer_relevancy", "faithfulness"],
        "impact_type": "Social_biases"
    }
}


for dataset_name, info in datasets_metrics.items():
    expset_name = f"{products}_base_on_{info['impact_type']}"
    expset_readme = f"Compliance Evaluation for {products} Product, based on {dataset_name} dataset, who analyze {info['impact_type']}"

    metrics = info["metrics"] + technical_metrics

    common_params = {
        "dataset": dataset_name,
        "model": {
            "extra_params": {"rag": {"mode": "rag", "limit": 7}},
            "sampling_params": {"temperature": 0.2},
        },
        "metrics": metrics,
        "judge_model": JUDGE,
    }
    grid_params = {
        "model": [
            {
                "name": "mistralai/Mistral-Small-3.2-24B-Instruct-2506",
                "aliased_name": "albert-large",
                "base_url": "https://api.albert.fr/v1", "api_key": os.getenv("ALBERT_API_KEY")
            },
            {
                "name": "meta-llama/Llama-3.1-8B-Instruct",
                "aliased_name": "albert-small",
                "base_url": "https://api.albert.fr/v1", "api_key": os.getenv("ALBERT_API_KEY")
            },
        ]
    }
    # Create the experiment set
    expset = {
        "name": expset_name,
        "readme": expset_readme,
        "cv": {"common_params": common_params, "grid_params": grid_params, "repeat": 3},
    }

    # Launch the experiment set
    response = requests.post(f"{EVALAP_API_URL}/experiment_set", json=expset, headers=headers)
    resp = response.json()
    if "id" in resp:
        print(f'Created expset: {resp["name"]} (ID: {resp["id"]})')
    else:
        print(f'Error creating experiment set for {dataset_name}: {resp}')

Viewing Experiment Results and Progress

After launching an experiment:

1. Navigate to the compliance details page : http://localhost:8501/compliance
2. View summary results showing:
   • Overall performance metrics for each model
   • Support table displaying the number of experiments used for score averaging
3. Explore detailed results:
   • Number of successful and failed attempts per experiment
   • Detailed results for each experiment

Next Steps:

Explore the notebooks for real examples of creating evaluations with EvalAP.