geh 0.2.0

CLI-first harness for safety and guardrail evaluation

CLI-first harness for benchmarking guardrail, moderation, and safety classification models.

Evaluate any safety model — local HuggingFace, vLLM, OpenAI, Anthropic, or custom API — against 80+ built-in safety benchmarks with a single command.

Quickstart

pip install geh

# Run a quick eval
geh run --dataset xstest --model mock --limit 50

# Run multiple datasets
geh run --dataset xstest,toxic_chat,harmful_qa --model hf \
    --model-name meta-llama/Llama-Guard-3-8B

# Run from a YAML config
geh run --config examples/run-mock-jsonl.yaml

# Use benchmark packs
geh run --pack core --model mock

Installation

Requires Python 3.10+.

# Base install
pip install geh

# With HuggingFace model support
pip install "geh[hf]"

# With vLLM support
pip install "geh[vllm]"

# With API model support (OpenAI, Anthropic)
pip install "geh[api]"

From source (for development):

git clone https://github.com/Virtue-Research/guard-eval-harness.git
cd guard-eval-harness
pip install -e ".[dev]"

Copy .env.example to .env and fill in the API keys you need.

Usage

Inline mode

The fastest way to run evals — no config files needed:

geh run --dataset <dataset> --model <adapter> [--model-name <name>] [options]

# HuggingFace model on XSTest
geh run --dataset xstest --model hf --model-name meta-llama/Llama-Guard-3-8B

# OpenAI moderation
geh run --dataset xstest,toxic_chat --model openai_moderation

# vLLM serving
geh run --dataset harmbench_behaviors --model vllm \
    --model-name meta-llama/Llama-Guard-3-8B --batch-size 32

# Limit samples for quick smoke tests
geh run --dataset xstest --model mock --limit 10

YAML config mode

For full control over model args, dataset options, execution tuning, and output:

geh run --config examples/run-mock-jsonl.yaml

See examples/ for sample configs.

Benchmark packs

Curated dataset bundles for common evaluation scenarios:

geh list packs
geh run --pack core --model mock
geh run --pack jailbreak --model hf --model-name meta-llama/Llama-Guard-3-8B

Discovery

geh list datasets    # 80+ built-in safety benchmarks
geh list backends    # Available model adapters
geh list packs       # Curated benchmark bundles
geh list metrics     # Supported metrics

Inspecting results

geh inspect --run-dir out/my-run       # View manifest, summary, artifacts
geh report --run-dir out/my-run        # Rebuild HTML report
geh compare --run-a out/run1 --run-b out/run2  # Diff two runs
geh export --run-dir out/my-run --format csv --output results.csv

Run artifacts

Each run writes a self-contained directory:

out/my-run/
  manifest.json              # Run metadata
  resolved-config.json       # Exact config snapshot
  summary.json               # Aggregated metrics
  report.html                # Static HTML report
  datasets/
    <dataset>/
      predictions.jsonl      # Per-sample predictions
      metrics.json           # Dataset-level metrics
      dataset-manifest.json  # Dataset metadata

Model adapters

Adapter	Description
mock	Deterministic mock for testing
hf	HuggingFace Transformers (local GPU)
vllm	vLLM inference server
openai_compatible	OpenAI-compatible APIs
openai_moderation	OpenAI Moderation endpoint
anthropic	Anthropic Claude API
http	Generic HTTP endpoint

Datasets

80+ built-in safety benchmarks spanning two modalities:

Text

The core modality — evaluate text-based guardrails and moderation models across a range of safety dimensions:

• Jailbreak / adversarial: XSTest, HarmBench, JBB Behaviors, AdvBench, Do-Anything-Now, StrongREJECT, MaliciousInstruct, WildGuardMix
• Toxicity: ToxicChat, ToxiGen, Jigsaw Toxicity, Civil Comments, RealToxicityPrompts, OR-Bench
• Hate & harassment: HateCheck, DynaHate, ETHOS, HatExplain, Implicit Hate, Measuring Hate Speech, Social Bias Frames, ConvAbuse
• General safety: BeaverTails 330k, Do-Not-Answer, OpenAI Moderation (via API), GuardBench, CircleGuardBench
• Prompt injection: Dedicated prompt-injection benchmarks for testing input-filtering guardrails

Image

Evaluate multimodal safety models that process image+text inputs. The harness handles image downloading, caching, and normalization automatically:

• Unsafe content detection: UnsafeBench (8k+ images across safety categories), HoliSafeBench (holistic image safety with fine-grained risk types)
• Visual jailbreaks: JailbreakV (adversarial images designed to bypass vision-language model safeguards)
• Image edit safety: Safe-vs-Unsafe Image Edits (detecting harmful image manipulation requests)
• Cross-modal attacks: VLSBench, MSTS (text+image multimodal safety evaluation)
• Benign baselines: ImageNet-1k safe subset (measuring false positive rates on benign images)
• Local image data: Load from local directories or JSONL manifests with image paths/URLs

Local files

Bring your own data in any modality:

• local_jsonl — text samples from a JSONL file
• local_csv — text samples from a CSV file
• local_image_jsonl — image+text samples from a JSONL manifest with image paths/URLs
• local_image_dir — image samples from a directory of images

Run geh list datasets for the full list.

Secure-coding agents

Beyond classification, the harness also runs repository-level secure-coding benchmarks under geh vibe: a coding agent writes or completes real code, and an out-of-process oracle builds it in a container to score functional correctness and security. See the VibeCoding Bench guide and geh vibe datasets.

About

guard-eval-harness is built and maintained by the research team at Virtue AI — one security solution for your entire AI stack.

License

MIT