Skip to content
ai-supply.store
DiscoverCategoriesLeaderboardsCommunityAgent APIFAQ
Sign inSign up free
catalog / Research / MMLU Benchmark
▣DatasetResearchFree

MMLU Benchmark

MMLU: 57-subject multiple-choice benchmark for broad LLM knowledge and reasoning.

@ai-supply
Installs44k
↗ Source repository
← More ResearchResearch leaderboard →How we grade security →Source ↗

MMLU — Measuring Massive Multitask Language Understanding

The official code and data for MMLU (ICLR 2021), a benchmark of roughly 16,000 multiple-choice questions spanning 57 subjects — from elementary math and US history to law, medicine, and computer science.

It measures a model's broad world knowledge and problem-solving ability in zero- and few-shot settings and remains one of the most widely reported LLM benchmarks.

MIT licensed and lightweight, making it easy to run and integrate into evaluation pipelines.

Rating rank
#1
of 17 in Research
Install rank
#11
of 17 in Research
Security score
100/100 · A
safe
Security rank
#1
of 17 in Research
Installs
44k
cat avg 51k
This listing vs category average
Installs
this
cat avg
Security (of 100)
this
cat avg
Adoption trend
See the Research leaderboard →
✓ Security: Safe · 100100/100 · grade Ascanned 14h ago
✓ no compromise signals3 risk-surface · 5/20 OWASP controls flagged

Compromise signals — malicious or tampered code (leaked secrets, backdoors, a dropped executable) — reduce the score, and known dependency CVEs carry a bounded penalty (they warrant review but never QUARANTINE — update the dependency to clear). Other dangerous-by-capability traits are risk surface, expected for some capabilities. Every finding is mapped to its OWASP control below.

Data card · low confidence (static)
license: detected
8 files

Findings mapped to the OWASP Top 10 for LLM Applications (2025) and the OWASP Machine Learning Security Top 10. Expand any flagged control for the exact findings — compromise reduces the score; expected/risk-surface do not.

OWASP Top 10 for LLM Applications
⚠LLM05Improper Output Handlingmedium
Code that pipes model/user output into shell, eval, SQL or paths unsafely.
•Suspicious code patterns — dynamic code execution · hendrycks-test-4450500/evaluate.py (CWE-95)risk surface
⚠LLM10Unbounded Consumptionmedium
Unbounded loops/recursion causing DoS or runaway cost.
Enforced at runtime by the gateway (rate limits + spend caps + size caps); static check flags unbounded loops.
•Potentially unbounded loop — an infinite loop (while True / while(1) / for(;;)) may cause runaway consumption · hendrycks-test-4450500/crop.py (CWE-835)risk surface
⚠LLM06Excessive Agencylow
Over-broad tool/permission surface or unrestricted egress.
•Broad capability surface — 3 high-impact capability categories referenced — verify least-privilege · hendrycks-test-4450500/crop.py (CWE-272)risk surface
§LLM09MisinformationGovernance
Artifacts designed to produce false/deceptive output.
Detectable only by runtime behavioral evaluation; addressed via responsible-use attestation.
✓LLM01Prompt InjectionPassed
✓LLM02Sensitive Information DisclosurePassed
✓LLM03Supply ChainPassed
✓LLM04Data and Model PoisoningPassed
Backdoors/poisoning in training data or serialized models.
Behavioral poisoning needs model execution; static check covers unsafe serialization + dataset skew only.
✓LLM07System Prompt LeakagePassed
✓LLM08Vector and Embedding WeaknessesPassed
PII or plaintext source leakage in embedding/vector exports.
Embedding inversion/poisoning is largely runtime; static check covers PII in vector exports.
OWASP Machine Learning Security Top 10
⚠ML09Output Integritymedium
Middleware tampering with model outputs in transit.
Gateway enforces TLS + response integrity; static check flags output-rewriting code.
•Suspicious code patterns — dynamic code execution · hendrycks-test-4450500/evaluate.py (CWE-95)risk surface
⚠ML05Model Theftlow
Unlicensed re-distribution / license-incompatible derivatives.
Static check verifies license declaration; extraction throttling is runtime.
•No license signal — no SPDX id or license keyword found · hendrycks-test-4450500/README.mdrisk surface
§ML01Input Manipulation (Adversarial)Governance
Models vulnerable to adversarial perturbations.
Requires runtime robustness evaluation; addressed via publisher robustness attestation.
§ML03Model InversionGovernance
Training data reconstructable from a model's outputs.
Runtime/evaluation property; addressed via model-card data-provenance + DP attestation.
§ML04Membership InferenceGovernance
Determining whether a record was in the training set.
Runtime/evaluation property; addressed via overfitting disclosure + DP attestation.
§ML08Model SkewingGovernance
Models trained on skewed data producing biased output.
Requires fairness evaluation; addressed via model-card bias/limitations disclosure.
✓ML02Data PoisoningPassed
Poisoned training datasets with triggers or anomalous distributions.
Static check covers trigger phrasing, PII and label skew; full poisoning detection is runtime.
✓ML06AI Supply ChainPassed
✓ML07Transfer Learning AttackPassed
Backdoored base models / LoRA adapters propagating to derivatives.
Backdoor detection needs behavioral probing; static check covers unsafe serialization + provenance.
✓ML10Model Poisoning (Weights)Passed
Tampered model weight files; integrity must be verifiable.
Static check enforces safe formats + records a content hash for downstream verification.
Other findings (1) · hygiene / uncategorized
•Unrecognized file type — '.?' is not on the allowlist · hendrycks-test-4450500/LICENSErisk surface
✔ verified source · pinned hendrycks-test-4450500
Check against a policy

The same gate an agent runs before installing (POST /api/v1/trust/mmlu-benchmark/check). Click a policy:

Consume MMLU Benchmark programmatically. Authenticate with an API key or session — see Authorize an agent.

# Agents: CHECK BEFORE YOU INSTALL (no auth) — score, grade, level, capability manifest
curl https://ai-supply.store/api/v1/trust/mmlu-benchmark

# Gate against your org policy (returns { pass, violations })
curl -X POST https://ai-supply.store/api/v1/trust/mmlu-benchmark/check \
  -H "Content-Type: application/json" \
  -d '{"minGrade":"B","denyPermissions":["shell"],"denyUnknownEgress":true}'

# CLI
npx ai-supply add mmlu-benchmark

# REST (install → download)
curl -X POST https://ai-supply.store/api/v1/listings/mmlu-benchmark/install \
  -H "Authorization: Bearer $AIM_KEY"

# MCP tool
install_listing({ "slug": "mmlu-benchmark" })
OpenAPI spec →
vlatest
✓ Security: Safe · 10015h ago

Curated mirror — latest upstream source. See the repository for tagged releases.

Sign in and install this listing to leave a review.

More from @ai-supply

View profile →
◉Agent
MetaGPT
Multi-agent framework that assigns GPT roles (PM, engineer, QA) to solve complex software tasks end-to-end.
↓ 1.0M
⇄Connector
vLLM
High-throughput, memory-efficient LLM inference engine with PagedAttention and continuous batching.
↓ 892k
⇄Connector
Meilisearch
Lightning-fast open-source search engine with typo-tolerance, semantic hybrid search, and sub-50ms response times.
↓ 811k
△Eval
Weights & Biases (wandb)
ML experiment tracking and visualization — log metrics, hyperparameters, models, and media in real time.
↓ 784k
ai-supply.store

Free, security-vetted AI capabilities — skills, MCPs, plugins, agents, datasets and more, each graded and freshness-tracked, and built for humans and agents alike.

api · v3.1status · all green
Contact
support@ai-supply.storesecurity@ai-supply.store
Catalog
  • Discover
  • Categories
  • Leaderboards
  • Benchmarks
  • Security
Community
  • Community
  • FAQ
For agents
  • Quickstart (60s)
  • Authorize an agent
  • Agent API
  • OpenAPI spec
For builders
  • Publish
  • Dashboard
Account
  • Create account
  • Sign in
  • Settings
Legal
  • Terms
  • Publisher Agreement
  • Acceptable Use
  • Privacy