README.md

metadata

pretty_name: 'MMLU: 500 random test questions (QiSpace)'
license: mit
tags:
  - lightrider
  - quantum-provenance
  - evaluation
  - subset
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-00000-of-00001.parquet

MMLU: 500 random test questions (QiSpace)

500 rows chosen at random from the 14042 rows of the test split of cais/mmlu (all), revision c30699e8356d. The rows are unchanged; _lr_source_index gives each row's position in the source split.

The rows keep the source dataset's MIT license (license text).

The dataset was generated from a seed drawn from qispace_kds_001 through Light Rider. The seed comes with a signed receipt, so anyone can check where the data came from and regenerate it exactly.

Provenance

Entropy source	qispace_kds_001
Drawn	2026-10-19 17:54 UTC
Light Rider request	lr_sreq_4e103f300d674c3c83421b3a2be5b913
Seed	9f51784980c33e204321de4d235ae41be73ab9a55980c898a7f0d9a69b543ec9
Generator	eval_subset/1
Rows	500

Parameters:

{
  "source_dataset": "cais/mmlu",
  "config": "all",
  "split": "test",
  "revision": "c30699e8356da336a370243923dbaf21066bb9fe",
  "source_num_rows": 14042,
  "n": 500,
  "requested": {
    "n": 500,
    "fraction": null
  }
}

Full technical record

Receipt signature	Ed25519, key fingerprint (SHA-256) 363b9750db0acd16ab188e4ecc8a58e4f811b28f5870cfc3ba4522274a11645e
Output commitment (signed)	sha256:8d81a1f45d1c1cfec1e7ffb372d89f7eb7eccfd3e5b0b5b061fac8cb8e2cb350
Commitment salt	6155d874ceb1462bbcba8a7db822419fe9b7b01019b872122022a0eb19246a44
Generator stream	LR-ChaCha20 v1, domain eval_subset
Content hash (lr-canonical-jsonl/1)	729f451c60bb7ff33073a161dd2316e668b5f1db6cf438d8aabf9ef82b932a53
Built with	lightrider-huggingface 0.1.1, lightrider 1.6.2, pyarrow 26.0.0

The complete record is in lightrider_provenance.json. The signed receipt, exactly as issued, is in receipt.json.

How to verify

pip install lightrider-huggingface

import lightrider_huggingface as lrhf

print(lrhf.verify("LightRiderInc/mmlu-random-sample-qispace"))

This needs no Light Rider account. It prints each check as PASS or FAIL.

For a subset of a very large dataset, lrhf.verify(..., indices_only=True) skips re-downloading the source.

How verification works

1. The receipt is genuine. Light Rider signs every draw's receipt (Ed25519). The public key is at entropy.lightriderinc.com/api/v1/public-key.
2. The seed is that draw's output. The signed receipt includes a hash of the exact bytes drawn, so a seed from any other draw doesn't match.
3. The data follows from the seed. Generation is deterministic, so verify rebuilds every row from the seed and parameters and compares it with the published data.

What this doesn't claim

That quantum randomness makes this dataset better than one built with a good classical generator. It's an alternative source whose origin you can check. A receipt also can't show that the publisher drew only once; its timestamp shows when the draw happened.