← MMLU Leaderboard

mmlu@2024-03 · Citation-only

The MMLU citation.

Fifty-seven academic subjects, five-shot. Every score in this section is copied from a public leaderboard — rendered so the record is complete, marked reported so it counts at the 1.0× lane only after the human gate.

Provenance: reported. Scores here count at the reported 1.0× Trust Magnitude lane after human gate approval. They are not Gaia CI/verifier reproduced.

01What is MMLU

Massive Multitask Language Understanding measures a language model's knowledge and reasoning across 57 academic subjects — mathematics, history, law, medicine, computer science — in a 5-shot setting. The model sees five example questions with answers before being tested on new ones.

Original paper: Measuring Massive Multitask Language Understanding (Hendrycks et al., 2020).

02Snapshot source

Scores in this registry are copied from a static snapshot of the HuggingFace Open LLM Leaderboard dated 2024-03-01. The snapshot file lives at scripts/benchmarks/mmlu/snapshot.json.

FieldValue
benchmarkIdmmlu@2024-03
unitpct (0..100 percentage accuracy)
sourceSnapshotDate2024-03-01
runAt2024-03-01T00:00:00Z

03Provenance ladder

The registry requires reproducible provenance for Trust Magnitude contribution. MMLU scores in this snapshot are cited from a public leaderboard — they were not produced by a CI-executed harness in this repository and have not been co-signed by a 4★+ Verifier running the model directly.

ProvenanceTMHow achieved
verified2.0×CI-reproduced or verifier-attested; legacy aliases ci-reproduced/verifier-attested
reported1.0×Public claim or mirrored leaderboard accepted by the human gate; legacy alias mirrored
rejectedBlacklisted, disputed, pending, candidate, retired, or unknown

A cited number can be stale. It can be measured under different prompts, different splits, a different tokenizer. Counting it would inflate the badge.

The leaderboard renders reported rows with a "Reported" badge to surface the distinction without hiding the data. This is deliberate: the record is complete, the score is legible, and TM stays clean.

04Refreshing the snapshot

  1. Visit the leaderboard.
  2. Export current 5-shot MMLU averages for the skills in snapshot.json.
  3. Edit scripts/benchmarks/mmlu/snapshot.json; bump sourceSnapshotDate.
  4. Run python scripts/benchmarks/mmlu/ingest.py --dry-run to preview.
  5. Run GAIA_OPERATOR_OVERRIDE=1 python scripts/benchmarks/mmlu/ingest.py to write.
  6. Regenerate docs/api/v1/benchmarks/mmlu.json.
  7. Open a PR on a review/meta/ branch.

05API projection

Machine-readable row data is served at /api/v1/benchmarks/mmlu.json. The index of all registered benchmarks lives at /api/v1/benchmarks/index.json.