sWuggy

sWuggy is a spoken lexical-discrimination benchmark: each item is a pair consisting of a real word and a phonotactically matched pseudo-word, synthesized as audio. A model is scored on whether it assigns higher probability to the real word than to its matched pseudo-word. This repository hosts the audio (as WebDataset tar shards) together with the per-item metadata needed to run the evaluation.

Evaluation only. This is a held-out benchmark for evaluating spoken language models. It must not be used as training data. Training on these items invalidates any sWuggy score and contaminates the benchmark for everyone.

Status: A companion repository with the evaluation code for this benchmark is a work in progress and will be linked here once released. For now this repository provides the data and the file layout described below.

Repository layout

Audio is packed into WebDataset tar shards (many small audio files would otherwise hit the Hub's 10,000-files-per- directory limit and make download/streaming slow). Metadata stays as plain CSV/TXT files alongside the shards.

inftrain/
  en/
    audio/         testset_{1,2,4,8,16,32,64}-{000000..NNNNNN}.tar   # audio shards
    frequencies/   testset_{1,2,4,8,16,32,64}.csv                    # per-item metadata
    gold.csv                                                          # gold labels
  fr/
    audio/         testset_{1,2,4,8,16,32,64}-{000000..NNNNNN}.tar
    frequencies/   testset_{1,2,4,8,16,32,64}.csv
    gold.csv
zrc2021/
  en/
    dev-{000000..NNNNNN}.tar                           # audio shards
    test-{000000..NNNNNN}.tar
    dev_filesmap.txt
    dev.gold.csv
    test_filesmap.txt
    test.gold.csv

There are two top-level collections:

• inftrain/ — the main sWuggy evaluation material, split by language (en, fr) and by test set (testset_1 … testset_64). Despite the directory name, this is evaluation data and is not for training.
• zrc2021/ — the ZeroSpeech 2021 sWuggy split (dev / test), under en/.

Inside a shard

Each tar shard is a plain POSIX tar. Every audio file is stored under its basename, so the WebDataset key is the filename stem. Files that share a key are grouped into the same example; here each example is a single audio file, and its label/metadata lives in the accompanying CSV keyed by the same filename. Audio is .ogg for inftrain/ and .wav for zrc2021/.

Metadata files

• gold.csv / *.gold.csv — gold labels per item.
• frequencies/testset_N.csv — per-item lexical frequency information.
• *_filesmap.txt (zrc2021) — mapping between item identifiers and audio files.

Accessing the data

Stream the audio with datasets

The audio is in WebDataset format, so it is loaded with the webdataset builder. Map a glob of shards to a split with data_files, and stream with streaming=True to avoid downloading everything up front:

from datasets import load_dataset

# One test set of the English inftrain material:
ds = load_dataset(
    "webdataset",
    data_files={"test": "hf://datasets/coml/sWuggy/inftrain/en/audio/testset_1-*.tar"},
    split="test",
    streaming=True,
)

for sample in ds:
    # The audio column is named after the file suffix inside the tar
    # ("ogg" for inftrain, "wav" for zrc2021); the key is in sample["__key__"].
    key = sample["__key__"]
    audio = sample["ogg"]   # bytes; decode with soundfile/librosa as needed
    break

Load several test sets at once by mapping each to its own split:

data_files = {
    f"testset_{n}": f"hf://datasets/coml/sWuggy/inftrain/en/audio/testset_{n}-*.tar"
    for n in (1, 2, 4, 8, 16, 32, 64)
}
ds = load_dataset("webdataset", data_files=data_files, streaming=True)

The ZeroSpeech 2021 split works the same way with the wav suffix:

ds = load_dataset(
    "webdataset",
    data_files={
        "dev":  "hf://datasets/coml/sWuggy/zrc2021/en/dev-*.tar",
        "test": "hf://datasets/coml/sWuggy/zrc2021/en/test-*.tar",
    },
    streaming=True,
)

Read the metadata

The CSV/TXT metadata can be read directly over hf:// without downloading the audio:

import polars as pl

gold = pl.read_csv("hf://datasets/coml/sWuggy/inftrain/en/gold.csv")
freqs = pl.read_csv("hf://datasets/coml/sWuggy/inftrain/en/frequencies/testset_1.csv")

Join metadata to audio on the WebDataset key (the filename stem) where you need both.

Download specific files

To pull individual shards or metadata files to local disk:

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="coml/sWuggy",
    repo_type="dataset",
    filename="inftrain/en/gold.csv",
)

Or snapshot a subset with huggingface_hub.snapshot_download(..., allow_patterns=...).

Evaluation

The evaluation expects, for each word / pseudo-word pair, a model score (typically a log-probability or pseudo log-likelihood) for both audio items; accuracy is the fraction of pairs where the real word is scored above its matched pseudo-word.

Evaluation code that runs this end-to-end is being prepared in a separate repository and will be linked here.

License

Released under the Academic Free License v3.0 (afl-3.0).

Citation

Contact

Questions and issues: please open a discussion in the Community tab of this repository.