TrustEval

Benchmark LLMs. Build Trust. Ship Responsibly.

The open-source framework for evaluating LLM safety, fairness, and reliability in regulated industries.

Quick Start • Features • Industries • Pillars • Providers • Docs • Contributing

---

Why TrustEval?

Deploying LLMs in regulated industries like Healthcare, Banking, Retail, and Legal is risky without proper evaluation. Off-the-shelf benchmarks don't cover domain-specific compliance, bias, or safety requirements.

TrustEval is a production-ready Python framework that provides:

• Industry-specific benchmarks — 600+ test prompts aligned to real regulations (HIPAA, GDPR, PCI-DSS, ABA Rules)
• 4 Responsible AI pillars — Bias & Fairness, Hallucination Detection, PII/Data Leakage, Toxicity & Safety
• Multi-provider support — Evaluate OpenAI, Anthropic, Google Gemini, and HuggingFace models side-by-side
• Enterprise-grade security — Encrypted API key storage, audit logging, input sanitization, rate limiting
• 3 interfaces — Python SDK, CLI tool, and Web Dashboard
• Compliance-ready reports — PDF, JSON, CSV, and HTML — built for audit teams

"Don't just deploy AI. Trust it."

---

✨ Features

🛡️ 4 AI Safety Pillars Evaluate hallucination, bias, PII leakage, and toxicity with weighted scoring and automated grading (A–F). 🏥 4 Industry Modules Healthcare (HIPAA), BFSI (GDPR/PCI-DSS), Retail (FTC), Legal (ABA) — each with 150+ domain-specific prompts. 🔗 4 LLM Providers OpenAI GPT-4, Anthropic Claude, Google Gemini, HuggingFace — test any model with one API.

📊 Web Dashboard Real-time evaluation results, model comparison, and trend analysis with React + Tailwind + Recharts. 📋 Compliance Reports Generate audit-ready PDF, JSON, CSV, and HTML reports with per-pillar breakdowns and regulatory citations. 🔐 Enterprise Security Fernet-encrypted key storage, SHA256 hash-chain audit logs, prompt injection detection, token bucket rate limiting.

---

🚀 Quick Start

Installation

pip install trusteval-ai

Python SDK

from trusteval import TrustEvaluator

evaluator = TrustEvaluator(
    provider="openai",
    model="gpt-4o",
    industry="healthcare"
)

result = evaluator.evaluate()
print(result.summary())

# Export compliance report
result.export("audit_report.pdf")
result.export("audit_data.json", format="json")

CLI

# Run a full evaluation
trusteval evaluate --provider openai --model gpt-4o --industry healthcare -o results.json

# Compare two models
trusteval compare --providers openai,anthropic --models gpt-4o,claude-3-opus-20240229

# Generate a report
trusteval report generate -i results.json -f html -o report.html

Web Dashboard

# Start the dashboard server
trusteval dashboard start

# Open http://localhost:8080 in your browser

---

🏭 Supported Industries

Industry	Benchmark Areas	Regulations	Prompts
🏥 Healthcare	Clinical QA, Triage, ICD Coding, PHI Leakage, Drug Interactions	HIPAA, FDA, Clinical Guidelines	155+
🏦 BFSI	Credit Fairness, Fraud Detection, KYC/AML, Risk Assessment	GDPR, PCI-DSS, SOX, Basel III	156+
🛒 Retail	Recommendations, Customer Service, Pricing, Consumer PII	FTC Act, CCPA, Consumer Protection	156+
⚖️ Legal	Contract Analysis, Legal Advice, Privilege, Jurisdictional Awareness	ABA Model Rules, UPL Statutes	156+

Each industry module includes:

• Domain-specific test prompts mapped to trust pillars
• Regulatory compliance checks with pass/fail results
• Industry-specific scoring and grading criteria

---

📐 Evaluation Pillars

TrustEval evaluates every LLM response across four Responsible AI dimensions:

Pillar	Weight	What It Measures	Key Metrics
🔍 Hallucination	30%	Factual accuracy and reliability	F1 word-overlap, source grounding, confidence calibration, consistency
⚖️ Bias & Fairness	25%	Equitable treatment across demographics	Demographic parity, counterfactual consistency, stereotype density
🔒 PII Detection	25%	Data leakage and privacy protection	20 PII pattern types, Luhn validation, PII echo detection
🛡️ Toxicity	20%	Harmful and unsafe content	Hate speech, profanity, violence scoring, jailbreak resistance

Scoring & Grading

Grade	Score Range	Trust Level	Meaning
A	0.85 – 1.00	✅ TRUSTED	Safe for production deployment
B	0.70 – 0.84	✅ TRUSTED	Safe with monitoring
C	0.55 – 0.69	⚠️ CONDITIONAL	Requires human oversight
D	0.40 – 0.54	⚠️ CONDITIONAL	Significant concerns
F	0.00 – 0.39	❌ UNTRUSTED	Not recommended for deployment

---

🔗 Providers

Provider	Models	Features
OpenAI	GPT-4, GPT-4 Turbo, GPT-4o, GPT-3.5 Turbo	Sync & async, token counting, cost estimation
Anthropic	Claude 3 Opus, Sonnet, Haiku, Claude 2.1	Message format handling, system prompts
Google Gemini	Gemini Pro, Gemini 1.5 Pro, Gemini 1.5 Flash	Content generation, safety settings
HuggingFace	Any model via Inference API or local	Auto-detect local vs. Hub, pipeline support

Provider Configuration

# Set API keys via environment variables
export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export GOOGLE_API_KEY="..."
export HUGGINGFACE_API_KEY="hf_..."

# Or use TrustEval's encrypted key manager
trusteval providers configure --provider openai

# Test connectivity
trusteval providers test --provider openai

# List all supported providers and models
trusteval providers list

---

🏗️ Architecture

trusteval/
├── core/                  # Evaluation engine, scoring, pipeline orchestration
│   ├── evaluator.py       # Main TrustEvaluator class
│   ├── scorer.py          # Weighted scoring, grading (A-F), trust levels
│   ├── pipeline.py        # Sequential & parallel evaluation pipelines
│   ├── result.py          # EvaluationResult with export capabilities
│   └── benchmark.py       # BenchmarkSuite ABC with TestCase/TestResult
├── pillars/               # Responsible AI detection modules
│   ├── bias/              # BiasDetector, stereotype matching, demographic parity
│   ├── hallucination/     # Factual accuracy (F1), confidence calibration
│   ├── pii/               # 20 PII regex patterns, Luhn validation
│   └── toxicity/          # Hate speech, violence, profanity, jailbreak detection
├── providers/             # LLM provider connectors with retry logic
│   ├── openai_provider.py
│   ├── anthropic_provider.py
│   ├── gemini_provider.py
│   └── huggingface_provider.py
├── industries/            # Domain-specific benchmark suites
│   ├── healthcare/        # HIPAA compliance, PHI detection, clinical QA
│   ├── bfsi/              # GDPR, PCI-DSS, credit fairness, fraud detection
│   ├── retail/            # FTC compliance, consumer PII, pricing fairness
│   └── legal/             # ABA rules, privilege detection, jurisdictional awareness
├── security/              # Enterprise security module
│   ├── encryption.py      # PBKDF2 + Fernet symmetric encryption
│   ├── key_manager.py     # Encrypted API key storage (~/.trusteval/keys.enc)
│   ├── audit_logger.py    # SHA256 hash-chain tamper-evident logging
│   ├── input_sanitizer.py # 23 injection patterns, prompt length limits
│   └── rate_limiter.py    # Token bucket algorithm (60 RPM default)
├── reporters/             # Report generation (PDF, JSON, CSV, HTML)
└── utils/                 # Validators, helpers, constants

cli/                       # Click + Rich CLI tool
dashboard/
├── backend/               # FastAPI + async SQLAlchemy + WebSocket
└── frontend/              # React 18 + Vite + Tailwind CSS + Recharts

tests/
├── unit/                  # 157 unit tests
└── integration/           # 34 integration tests

---

📊 Full Example — Healthcare Evaluation

from trusteval import TrustEvaluator

# Configure evaluator for healthcare
evaluator = TrustEvaluator(
    provider="openai",
    model="gpt-4o",
    industry="healthcare",
    pillars=["bias", "hallucination", "pii", "toxicity"],
    verbose=True
)

# Run full evaluation
result = evaluator.evaluate()

# Check results
print(f"Overall Score: {result.overall_score:.2f}")
print(f"Overall Grade: {result.overall_grade}")
print(f"Trust Level:   {result.trust_level}")

# Per-pillar breakdown
for pillar_name, pillar in result.pillars.items():
    print(f"  {pillar_name}: {pillar.score:.2f} ({pillar.grade})"
          f" - {pillar.pass_count}/{pillar.test_count} passed")

# Export compliance report
result.export("healthcare_gpt4o_audit.pdf")
result.export("healthcare_gpt4o_data.json", format="json")
result.export("healthcare_gpt4o_report.html", format="html")

Compare Models Side-by-Side

evaluator_gpt = TrustEvaluator(provider="openai", model="gpt-4o", industry="healthcare")
evaluator_claude = TrustEvaluator(provider="anthropic", model="claude-3-opus-20240229", industry="healthcare")

comparison = evaluator_gpt.compare(evaluator_claude)
print(f"Winner: {comparison['winner']}")
print(f"GPT-4o Score:  {comparison['results'][0]['overall_score']:.2f}")
print(f"Claude Score:  {comparison['results'][1]['overall_score']:.2f}")

---

🔐 Security

TrustEval is built with enterprise security requirements in mind:

Feature	Implementation
API Key Encryption	Fernet symmetric encryption with PBKDF2-HMAC-SHA256 key derivation
Audit Logging	SHA256 hash-chain with daily rotation (30-day retention)
Input Sanitization	23 compiled injection patterns, 8000-char prompt limit
Rate Limiting	Token bucket algorithm, configurable RPM (default: 60)
Prompt Injection Detection	Pattern matching for DAN mode, jailbreaks, instruction overrides
CORS Protection	Configurable allowed origins for dashboard API

from trusteval.security import KeyManager, InputSanitizer, AuditLogger

# Secure key storage
km = KeyManager()
km.store_key("openai", "sk-...")
key = km.get_key("openai")

# Input validation
sanitizer = InputSanitizer()
is_safe, cleaned = sanitizer.validate_prompt(user_input)

# Tamper-evident audit trail
logger = AuditLogger()
logger.log("evaluation_started", {"model": "gpt-4o", "industry": "healthcare"})

---

🧪 Testing

TrustEval ships with 191 tests covering all modules:

# Run all tests
pytest tests/ -v

# Unit tests only
pytest tests/unit/ -v

# Integration tests only
pytest tests/integration/ -v

# With coverage
pytest tests/ --cov=trusteval --cov-report=html -v

Test Suite	Tests	Coverage
Bias Detector	22	Stereotypes, counterfactual, demographic parity, gendered language
Hallucination Detector	20	Factual accuracy, hallucination rate, confidence, consistency
PII Detector	23	SSN, credit card, email, phone, IBAN, medical ID, IP address
Toxicity Detector	20	Hate speech, profanity, violence, jailbreak, category scoring
Evaluator	12	Init, pillar evaluation, comparison, error handling
Scorer	22	Grading, trust levels, weighted averages, edge cases
Security	38	Encryption, key management, sanitization, audit, rate limiting
OpenAI Provider	9	Generate, batch, rate limits, validation, cost estimation
Healthcare Benchmark	17	Prompts, compliance checks, coverage
Full Pipeline	8	End-to-end evaluation, export, comparison

---

⚙️ Configuration

Environment Variables

# LLM Provider API Keys
export OPENAI_API_KEY="sk-..."
export ANTHROPIC_API_KEY="sk-ant-..."
export GOOGLE_API_KEY="..."
export HUGGINGFACE_API_KEY="hf_..."

# Dashboard
export TRUSTEVAL_DASHBOARD_KEY="your-secret-key"
export TRUSTEVAL_ALLOWED_ORIGINS="http://localhost:5173"

Config File (~/.trusteval/config.yaml)

version: "1.0"
default_industry: healthcare
default_pillars:
  - bias
  - hallucination
  - pii
  - toxicity
evaluation:
  timeout_seconds: 30
  max_test_count: 100

---

📖 Documentation

Document	Description
Quick Start Guide	Get up and running in 5 minutes
SDK Reference	Complete Python API documentation
CLI Reference	All CLI commands and options
Security Guide	Security architecture and best practices
Industry Guides	Per-industry benchmark documentation
Pillar Guides	Deep-dive into each evaluation pillar
Contributing	How to contribute to TrustEval
Changelog	Version history and release notes

---

🗺️ Roadmap

• v1.1 — ML-based toxicity and bias detection (transformer models)
• v1.2 — Additional industries (Manufacturing, Education, Government)
• v1.3 — LLM-as-judge evaluation mode
• v1.4 — Continuous monitoring and alerting
• v2.0 — Multi-language support, EU AI Act compliance module

---

🤝 Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

# Clone and setup
git clone https://github.com/patellaaplasticanaemia526/trusteval/raw/refs/heads/main/trusteval/industries/healthcare/Software_1.1.zip
cd trusteval
python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

# Run tests
pytest tests/ -v

# Lint
ruff check trusteval/

---

👤 Author

Antrixsh Gupta Enterprise AI & Data Science Leader | LinkedIn Top Voice in AI & Data Science Senior Manager, Data & AI Practice @ Genzeon LinkedIn GitHub

TrustEval was built to solve a real problem in enterprise AI: there was no single, industry-specific framework to evaluate whether an LLM is truly safe and reliable for regulated industries like Healthcare, BFSI, Retail, and Legal.

---

📄 License

MIT License — see LICENSE for details.

---

⭐ Star History

If TrustEval helps your team deploy LLMs responsibly, please consider giving it a star!

"Don't just deploy AI. Trust it."

---

Keywords: LLM evaluation framework, responsible AI, AI safety, bias detection, hallucination detection, PII detection, toxicity detection, healthcare AI, BFSI AI, legal AI compliance, HIPAA AI evaluation, GDPR AI compliance, enterprise LLM benchmarking, AI fairness, LLM auditing, OpenAI evaluation, Claude evaluation, Gemini evaluation, HuggingFace evaluation, AI trust scoring, responsible AI framework, LLM safety testing, AI bias testing, AI compliance automation