FairEval: Human-Aligned Evaluation for Generative Models

Author: Kriti Behl
GitHub: https://github.com/kritibehl/FairEval
Paper (preprint): “FairEval: Human-Aligned Evaluation for Generative Models”

FairEval is a lightweight research framework for evaluating LLM outputs beyond accuracy — focusing on:

• LLM-as-Judge alignment scoring
• Toxicity / safety analysis
• Human agreement metrics (κ, ρ)
• Group-wise fairness dashboards

It is designed as a research tool, not a deployment model.

---

What this repo contains

This Hugging Face repo currently serves as a model card + metadata hub for:

• The FairEval evaluation pipeline (code on GitHub)
• A planned Hugging Face Space demo (UI built in Streamlit)
• Links to my preprint and Medium explainer.

Code: https://github.com/kritibehl/FairEval
Medium: https://medium.com/@kriti0608/faireval-a-human-aligned-evaluation-framework-for-generative-models-d822bfd5c99d

---

Capabilities

FairEval supports:

1. Rubric-based LLM-as-Judge scoring

   • Uses a structured rubric (config/prompts/judge_rubric.md) to score:
     • coherence
     • helpfulness
     • factuality
   • Returns scalar scores that correlate with human preference.

2. Toxicity and safety metrics

   • Wraps a toxicity model (e.g., Detoxify) to compute:
     • composite toxicity
     • per-category scores (insult, threat, identity attack, etc.)
   • Provides Altair charts for:
     • toxicity breakdown by category
     • toxicity distribution by demographic group

3. Human evaluation agreement

   • Ingests a human_eval.csv file with human ratings.
   • Computes:
     • Fleiss’ κ (inter-rater reliability)
     • Spearman ρ between judge and human scores.

---

Example Usage

Checkout the GitHub repo and run the Streamlit demo:

git clone https://github.com/kritibehl/FairEval.git
cd FairEval
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
streamlit run demo/app.py