Skip to content

Benchmarks

28 of EvalGuard's 34 benchmark suites detailed here — 17 academic knowledge + reasoning suites and 11 safety/adversarial suites. Each ships a runnable case bank, a deterministic scorer, and reproducible runs.

Every case on this page was written by EvalGuard. These suites reproduce the task format and category taxonomy of the public benchmarks they are named after; they do not bundle those datasets. A suite named MMLU ships 26 EvalGuard-written questions, not the 14,042-question public MMLU test set, so an accuracy from it is a score on our questions and is not comparable to a published leaderboard or a model card. Where a real public split can be loaded over the network we say so on the suite and it runs under --from-hf; otherwise only the authored bank exists. Full per-dataset licensing is in DATASET-MODEL-LICENSES.md.
28
Detailed here
17
Academic
11
Safety / adversarial

Why benchmark coverage matters

A named suite gives a regression a shape: “faithfulness dropped” is an argument, “MMLU-format knowledge accuracy fell 6 points between these two prompt versions” is a number two engineers can act on. EvalGuard covers the seven task formats DeepEval markets (MMLU, HellaSwag, BIG-Bench Hard, DROP, TruthfulQA, HumanEval, GSM8K) plus 16 more academic suites — including medical, legal, finance, security and multimodal — and 11 safety suites. What that buys is comparability across your own runs: same questions, same scorer, same seed, every time. It does not buy comparability with a published leaderboard, because the cases are ours, not the public datasets'. Run any suite with one command:

evalguard benchmark run mmlu --model gpt-4o
evalguard benchmark run truthfulqa --model claude-opus-4
evalguard benchmark run humaneval --model gemini-2.5-pro

Academic & reasoning benchmarks (17)

Knowledge, reasoning, code, math, and vertical-domain suites for comparing models against each other and against your own earlier runs. Each reproduces the task format of the public benchmark it is named after; the cases are EvalGuard-written, and the count on each card is the count that suite ships.

MMLU

mmlu

General knowledge

Massive Multitask Language Understanding — 57 academic subjects from elementary school through professional-level (US history, computer science, law, medicine, ethics, etc.).

Cases: 26, written by EvalGuard
Task format after: Hendrycks et al., 2021 (ICLR)
Measures: Knowledge breadth across STEM, humanities, social sciences, professional domains.

BIG-Bench Hard (BBH)

bigbench-hard

Reasoning

Multi-step reasoning tasks in the BIG-Bench Hard style — the family where the prior-best LM averaged below the average human rater.

Cases: 20, written by EvalGuard
Task format after: Suzgun et al., 2023 (Google + Stanford)
Measures: Multi-step reasoning, symbol manipulation, logical deduction.

DROP

drop

Reading comprehension

Discrete Reasoning Over Paragraphs — requires multi-step arithmetic, counting, sorting over text-extracted facts.

Cases: 20, written by EvalGuard
Task format after: Dua et al., 2019 (NAACL)
Measures: Reading comprehension + numerical reasoning.

BoolQ

boolq

Reading comprehension

Yes/no question-passage pairs in the BoolQ style, drawn from natural search-query phrasing.

Cases: 25, written by EvalGuard
Task format after: Clark et al., 2019 (NAACL)
Measures: Reading comprehension for binary entailment.

TruthfulQA

truthfulqa

Factuality

Questions engineered to elicit imitative falsehoods — urban legends, common misconceptions, conspiracy theories.

Cases: 20, written by EvalGuard
Task format after: Lin et al., 2022 (ACL)
Measures: Truthfulness vs imitative falsehoods. Critical pre-deployment screen.

HellaSwag

hellaswag

Commonsense reasoning

Multiple-choice sentence completion testing commonsense reasoning. Designed adversarially so humans score 95%+ but models often fail.

Cases: 20, written by EvalGuard
Task format after: Zellers et al., 2019 (ACL)
Measures: Commonsense world understanding via adversarial completion.

HumanEval

humaneval

Code

Hand-written Python programming problems in the HumanEval style; pass@1 measured by running unit tests against generated code.

Cases: 20, written by EvalGuard
Task format after: Chen et al., 2021 (OpenAI Codex paper)
Measures: Code synthesis with executable verification.

GSM8K

gsm8k

Math

Grade-school math word problems requiring 2-8 step reasoning chains.

Cases: 20, written by EvalGuard
Task format after: Cobbe et al., 2021 (OpenAI)
Measures: Multi-step arithmetic reasoning, chain-of-thought quality.

ARC

arc

Science reasoning

Grade-school science questions in the ARC style, spanning the Easy and Challenge difficulty split.

Cases: 25, written by EvalGuard
Task format after: Clark et al., 2018 (AI2)
Measures: Science knowledge + multi-step reasoning.

BBQ

bbq

Bias

Bias Benchmark for QA — measures social biases in 9 demographic categories (age, disability, gender identity, nationality, physical appearance, race/ethnicity, religion, SES, sexual orientation).

Cases: 25, written by EvalGuard
Task format after: Parrish et al., 2022 (ACL)
Measures: Social bias under ambiguous + disambiguating contexts.

IFEval

ifeval

Instruction following

Verifiable instruction-following — measures whether a model follows specific formatting / length / structure constraints in a prompt.

Cases: 25, written by EvalGuard
Task format after: Zhou et al., 2023 (Google)
Measures: Programmatic verifiable instruction adherence.

MMMU

mmmu

Multimodal

Multi-discipline questions requiring combined image + text reasoning, in the MMMU style.

Cases: 31, written by EvalGuard
Task format after: Yue et al., 2024 (CVPR)
Measures: Vision-language reasoning at college level.

VisionBench

visionbench

Multimodal

Curated suite covering image-question reasoning, OCR, chart interpretation, scientific figure comprehension.

Cases: 30, written by EvalGuard
Task format after: EvalGuard-curated, drawn from MMMU + MMVet + ChartQA
Measures: Production vision-language model quality.

MedQA

medqa

Vertical — Medical

USMLE-style medical questions covering disease diagnosis, treatment and clinical ethics.

Cases: 31, written by EvalGuard
Task format after: Jin et al., 2021
Measures: Medical knowledge to USMLE board-passing standard.

LegalBench

legalbench

Vertical — Legal

Legal-reasoning tasks in the LegalBench style: issue spotting, rule recall, application, and conclusions.

Cases: 30, written by EvalGuard
Task format after: Guha et al., 2023 (NeurIPS Datasets)
Measures: Legal-domain reasoning at attorney-grade fidelity.

Financial Reasoning Benchmark

financebench

Vertical — Finance

Financial reasoning over calculations, accounting, market analysis, risk, and compliance. EvalGuard-authored synthetic corpus; the external FinanceBench dataset is not bundled.

Cases: 32, written by EvalGuard
Task format after: EvalGuard synthetic (bring-your-own-license)
Measures: Financial-domain reasoning.

CyberBench

cyberbench

Vertical — Security

Cybersecurity question bank across pentesting, vulnerability classification, threat intelligence, CWE/CVE matching.

Cases: 31, written by EvalGuard
Task format after: Liu et al., 2024
Measures: Cybersecurity domain knowledge.

Safety & adversarial benchmarks (11)

Adversarial datasets for measuring refusal quality, jailbreak resistance, and harm-category coverage. Each is referenced by ID in our red-team plugin registry (see /docs/plugins) so they double as both standalone benchmarks AND first-class red-team plugins.

AEGIS

aegis

Content safety

Content-safety probes across 13 risk categories. EvalGuard-authored synthetic corpus, following the AEGIS risk taxonomy — the dataset itself is not bundled.

Cases: 110, written by EvalGuard
Task format after: NVIDIA, 2024 (CC-BY-4.0)
Measures: Adversarial safety refusal on enterprise-grade risk taxonomy.

Harmful-Content Refusal

beavertails

Harmful content

Refusal probes across 14 harm categories. EvalGuard-authored synthetic corpus, inspired by the BeaverTails taxonomy — the dataset itself is not bundled.

Cases: 70, written by EvalGuard
Task format after: EvalGuard synthetic (BeaverTails taxonomy, bring-your-own-license)
Measures: Refusal across a broad harm taxonomy.

HarmBench

harmbench

Adversarial

Standardized harmful-behavior probes across 7 categories. EvalGuard-authored synthetic corpus, following the HarmBench behaviour taxonomy — the dataset itself is not bundled.

Cases: 100, written by EvalGuard
Task format after: Mazeika et al., CAIS 2024 (MIT)
Measures: Standardized adversarial defence baseline.

Pliny / L1B3RT4S

pliny

Jailbreak

Jailbreak-structure attacks — persona shift, chain-of-thought hijack, system-prompt extraction, reflective jailbreak. EvalGuard-authored re-implementations of the four canonical Pliny families; the L1B3RT4S archive is deliberately not bundled.

Cases: 50, written by EvalGuard
Task format after: elder-plinius, ongoing (MIT)
Measures: Real-world jailbreak resistance.

Toxic-Conversation Robustness

toxicchat

Production toxicity

Toxic-conversation robustness probes modeling production-distribution attack patterns. EvalGuard-authored synthetic corpus, inspired by the ToxicChat taxonomy — the dataset itself is not bundled.

Cases: 100, written by EvalGuard
Task format after: EvalGuard synthetic (ToxicChat taxonomy, bring-your-own-license)
Measures: Production-distribution attack patterns.

CyberSecEval

cyberseceval

Code security

Secure-code-generation and cyber-assistance-refusal probes spanning common CWEs and MITRE ATT&CK categories. EvalGuard-authored synthetic corpus, following the CyberSecEval taxonomy — the dataset itself is not bundled.

Cases: 100, written by EvalGuard
Task format after: Meta, 2024 (MIT)
Measures: Secure-code generation + cyber-attack assistance refusal.

Unsafe-Content Safety

unsafebench

Multimodal safety

Unsafe-content safety probes across 11 categories. EvalGuard-authored synthetic corpus, inspired by the UnsafeBench taxonomy — the dataset itself is not bundled.

Cases: 52, written by EvalGuard
Task format after: EvalGuard synthetic (UnsafeBench taxonomy, bring-your-own-license)
Measures: Vision-language safety alignment.

VLGuard

vlguard

Multimodal safety

Text re-grounding of the VLGuard safety taxonomy — privacy, deception, discrimination, dangerous actions — so a text-only firewall is exercised on the same harm categories. EvalGuard-authored.

Cases: 51, written by EvalGuard
Task format after: Edinburgh, 2024 (CC-BY-4.0)
Measures: Vision-language model fine-tuning + evaluation.

VLSU

vlsu

Multimodal safety

Text re-grounding of the VLSU underrepresented-harm categories: cultural erasure, minority-targeted misinformation, low-resource-language jailbreaks, disability stereotyping. EvalGuard-authored.

Cases: 50, written by EvalGuard
Task format after: Academic benchmark, 2024
Measures: VLM refusal calibration depth.

Do-Not-Answer

donotanswer

Refusal calibration

Prompts an LLM should decline or carefully caveat, across the Do-Not-Answer risk areas. EvalGuard-authored synthetic corpus — the LibrAI dataset itself is not bundled.

Cases: 52, written by EvalGuard
Task format after: LibrAI, 2023 (Apache 2.0)
Measures: Refusal calibration — too lax vs too strict.

XSTest

xstest

Over-refusal

Safe prompts that LOOK harmful, plus unsafe contrasts. EvalGuard-authored synthetic corpus, following the XSTest exaggerated-safety design — the dataset itself is not bundled.

Cases: 100, written by EvalGuard
Task format after: Bocconi, 2024 (CC-BY-4.0)
Measures: Over-refusal — the inverse failure mode to jailbreaks.

Public safety corpus — via the firewall-benchmark API

Apache-2.0 · public sample split · 10 source taxonomies

A unified, license-clean slice of the safety corpus is served from the EvalGuard firewall-benchmark API so the research community can benchmark refusal calibration, jailbreak resistance, and content-policy behaviour against the same prompts EvalGuard uses internally. The public sample split is fully visible; the held-out split is served as metadata only (category list + counts).

GET /api/v1/firewall-benchmark/public-corpus

Every prompt in the public corpus is original EvalGuard staff writing, inspired by 10 published safety-research taxonomies (AEGIS, BeaverTails, CyberSecEval, DoNotAnswer, HarmBench, ToxicChat, UnsafeBench, VLGuard, VLSU, XSTest). We license-audited the source corpora before release — the Pliny / L1B3RT4S jailbreak archive is intentionally not included.

Fetch it over HTTP (no API key required):

curl https://evalguard.ai/api/v1/firewall-benchmark/public-corpus

# Response: { meta, publicCorpus: [{ input, category, source, expectedBehavior }], note }
# For each prompt, check whether your model's response matches expectedBehavior
# ("block" for prompts that should be refused, "allow" for safe prompts).

CLI + SDK

Every benchmark is runnable from the CLI:

# Run a single benchmark (suite is a positional arg)
evalguard benchmark run mmlu --model gpt-4o

# Compare models — run the same benchmark once per model
evalguard benchmark run humaneval --model gpt-4o
evalguard benchmark run humaneval --model claude-opus-4

# Save the report (emit JSON with --json and redirect stdout)
evalguard benchmark run gsm8k --model gpt-4o --json > report.json

The CLI runs the suite locally (via @evalguard/core), so it needs no API key. The cross-model leaderboard — aggregated from eval and security-scan data — is also available from the Python SDK:

from evalguard import EvalGuardClient

client = EvalGuardClient(api_key="eg_live_...")
board = client.get_leaderboard(category="overall")
print(board["leaderboard"])  # [{"model": "gpt-4o", "overallScore": 0.87, ...}, ...]