# hle-glm53-20260923

Public raw data package for this measurement. Every number on the measurements page citing this window comes from these files.

Humanity's Last Exam, text only, 1,904 items, K=1: the 2026-09-18–19 suite at 65,536 output tokens, with truncated items re-run at 163,840. The dataset requests that questions and answers not enter corpora: `hle_sanitizado.json` does not include extracted answer or ground truth, only the per-item verdict and non-convergence flags. Full-file sha256 in the file itself.

Without judge (`judge_ok` null on all items): exact-match is a floor, not comparable to official score. In the earlier record, 45 items stored as `modo=infra` (re-run at 163,840 tokens and still unfinished; 1 multiple choice, 44 exact match) were left out of the denominators. They now count as not converged, that is, wrong: 453 multiple choice and 1,451 exact match. The type of each item comes from the dataset (`tipo_no_conjunto`). The published exact-match figure is strict: 127/1,451 = 8.75%; `exact.pct` (9.03%) still includes 4 answers accepted only by a relative tolerance and is retracted.

| file | what it is | sha256 (first 16) |
|---|---|---|
| `hle_sanitizado.json` | items per question or task, without question or response text; file lists what was removed | `f01d906d093fc5b5` |
| `resumo.json` | means, standard deviations, and window result (K repetitions per arm) | `87363b31f07d7266` |