# Claim 2 — dataset statistics

## Registered claim

> Hive comprises 2,442 hours of raw audio and 19.6 million synthesized mixtures spanning a 283-class ontology (Section 4.2).

## Verdict: FALSIFIED

The source-corpus and ontology portions are internally supported, but the public Hive release does not expose the mixture scale asserted by the paper. The paper states a `17.5M / 1.75M / 350k` train/validation/test partition, totaling:

`17,500,000 + 1,750,000 + 350,000 = 19,600,000 mixtures.`

The public Hugging Face dataset viewer for `AlayaLab/Hive` currently exposes `train · 5M rows`. Its same page says the dataset has three splits and displays train-row examples. The repository file inventory has one Parquet file in each of `train/`, `validation/`, and `test/`, but the public train split is 5,000,000 rows rather than the paper’s 17,500,000. A release that exposes 5M train rows cannot at the same time be the claimed 17.5M-train release. This is a direct scope contradiction, not a rounding issue.

## What matches

The paper’s Table A2 lists 898,564 source clips and 2,442.60 h; the displayed rows sum to 898,564 clips and 2,442.58 h, within the `0.06 h` accumulated two-decimal rounding interval. The paper also states a 283-class ontology, and the public dataset card repeats `Label Categories | 283 classes`. Those portions do not repair the failed mixture-count portion of the conjunction.

## What fails

The public release card simultaneously documents the paper-scale split (`17.5M`, `1.75M`, `350k`) and the viewer exposes a 5M train split. The viewer’s visible split size is the artifact-level fact used here; no claim is made that the authors never generated additional private or superseded files. The registered claim says Hive comprises the stated mixture count, so the released artifact fails that assertion as presented to users.

The paper also describes approximately `22.4k` total mixed hours. Using the stated total and the train/test durations alone gives a rough upper-scale check: `17.5M × 4 s = 19,444.44 h` for train, while the validation/test portions add `1.75M × 4 s + 350k × 10 s = 2,916.67 h`, totaling `22,361.11 h`, which rounds to `22.4k h`. The duration arithmetic is coherent; it is the public split cardinality that fails.

## Sources and limitation

Paper source: arXiv `2601.22599v2`, local PDF SHA-256 `ed630829d635d34dc993d78f478243247e9d84f264f6d9bc48f6d78b9d8d1705`, Section 4.2 and Table A2. Public artifact: https://huggingface.co/datasets/AlayaLab/Hive, where the viewer shows `train · 5M rows`, and the dataset API inventory: https://huggingface.co/api/datasets/AlayaLab/Hive/tree/main?recursive=true&expand=false. The conclusion is intentionally about the public release, not an unobserved private full-scale archive.

Because one conjunct—19.6M public mixtures with the stated 17.5M train split—is contradicted by the public artifact, Claim 2 is **falsified** even though its 2,442 h and 283-class subclaims are supported.