discophon 0.0.3__tar.gz

discophon-0.0.3/.github/workflows/release.yml

@@ -0,0 +1,109 @@
+name: Release Workflow
+permissions: {}
+
+on:
+  push:
+    tags:
+      - "*"
+
+jobs:
+  lint:
+    name: Linters
+    runs-on: ubuntu-latest
+    permissions:
+      security-events: write
+      contents: read # Fetch repository
+      actions: read
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
+        with:
+          persist-credentials: false
+
+      # - name: zizmor
+      #   uses: zizmorcore/zizmor-action@e639db99335bc9038abc0e066dfcd72e23d26fb4 # v0.3.0
+
+      - name: typos
+        uses: crate-ci/typos@bb4666ad77b539a6b4ce4eda7ebb6de553704021 # v1.42.0
+
+      - name: Ruff
+        uses: astral-sh/ruff-action@57714a7c8a2e59f32539362ba31877a1957dded1 # v3.5.1
+
+  build:
+    name: Build distribution
+    runs-on: ubuntu-latest
+    needs: lint
+    permissions:
+      contents: read # Fetch repository
+      actions: write # Upload artifact
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
+        with:
+          persist-credentials: false
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@61cb8a9741eeb8a550a1b8544337180c0fc8476b # v7.2.0
+        with:
+          version: "latest"
+          enable-cache: false
+
+      - name: Build wheel
+        run: uv build
+
+      - name: Upload dists
+        uses: actions/upload-artifact@b7c566a772e6b6bfb58ed0dc250532a479d7789f # v6.0.0
+        with:
+          name: python-package-distributions
+          path: dist/
+
+  publish-to-pypi:
+    name: Publish to PyPI
+    runs-on: ubuntu-latest
+    needs: build
+    environment:
+      name: publish-to-pypi
+    permissions:
+      actions: read # Download artifact
+      id-token: write # Needed for trusted publishing
+    steps:
+      - name: Download all dists
+        uses: actions/download-artifact@37930b1c2abaa49bbe596cd826c3c89aef350131 # v7.0.0
+        with:
+          name: python-package-distributions
+          path: dist/
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@61cb8a9741eeb8a550a1b8544337180c0fc8476b # v7.2.0
+        with:
+          version: "latest"
+          enable-cache: false
+
+      - name: Upload to PyPI
+        # run: uv publish --trusted-publishing always
+        run: uv publish -t ${{ secrets.PYPI_TOKEN }}
+
+
+  github-release:
+    name: GitHub release
+    runs-on: ubuntu-latest
+    needs: build
+    permissions:
+      actions: read # Download artifact
+      contents: write # Create GitHub release
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@8e8c483db84b4bee98b60c0593521ed34d9990e8 # v6.0.1
+        with:
+          persist-credentials: false
+
+      - name: Download all dists
+        uses: actions/download-artifact@37930b1c2abaa49bbe596cd826c3c89aef350131 # v7.0.0
+        with:
+          name: python-package-distributions
+          path: ./dist/
+
+      - name: Create release
+        env:
+          GITHUB_TOKEN: ${{ github.token }}
+        run: gh release create "${GITHUB_REF#refs/tags/}" ./dist/*

discophon-0.0.3/.gitignore

@@ -0,0 +1,7 @@
+__pycache__
+*.py[oc]
+build/
+dist/
+*.egg-info
+.venv
+tipa-mapping-*.pdf

discophon-0.0.3/.justfile

@@ -0,0 +1,17 @@
+set quiet
+
+[private]
+@default:
+    just --list
+
+# Test
+test:
+    uv run --group test pytest
+
+# Lint and format
+lint:
+    uv run --dev ruff check
+    uv run --dev ruff format
+    uv run --dev tombi format
+    uv run --dev typos
+    uv run --dev ty check src

discophon-0.0.3/.python-version

@@ -0,0 +1 @@
+3.12

discophon-0.0.3/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 CoML
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

discophon-0.0.3/PKG-INFO

@@ -0,0 +1,83 @@
+Metadata-Version: 2.4
+Name: discophon
+Version: 0.0.3
+Summary: The Phoneme Discovery Benchmark
+Author-email: CoML <dev@cognitive-ml.fr>
+License-Expression: MIT
+License-File: LICENSE
+Requires-Python: >=3.12
+Requires-Dist: filelock>=3.20.2
+Requires-Dist: httpx>=0.28.1
+Requires-Dist: joblib>=1.5.3
+Requires-Dist: numba>=0.63.1
+Requires-Dist: numpy>=2.3.5
+Requires-Dist: polars>=1.36.1
+Requires-Dist: praat-textgrids>=1.4.0
+Requires-Dist: soundfile>=0.13.1
+Requires-Dist: soxr>=1.0.0
+Requires-Dist: tqdm>=4.67.1
+Requires-Dist: xarray>=2025.12.0
+Provides-Extra: abx
+Requires-Dist: fastabx>=0.6.1; extra == 'abx'
+Description-Content-Type: text/markdown
+
+# The Phoneme Discovery benchmark
+
+[💾 [Website](https://benchmarks.cognitive-ml.fr/phoneme_discovery)] [📜 [Paper]()] [📖 [BibTex](https://github.com/bootphon/phoneme_discovery?tab=readme-ov-file#citation)]
+
+## Introduction
+
+The last several years have seen revolutionary improvements in both speech processing and textual natural language
+processing. In both cases, unsupervised or self-supervised pre-training has been the key to models autonomously
+discovering representations that are tremendously useful for doing language tasks. Yet, central to the study of human
+speech processing is the phoneme inventory, a small set of discrete units that abstract away from massive pronunciation
+variability in the signal.
+
+Discovering the correct set of phonemes for a language is crucial: encode the wrong categories, and contrasts between
+words are distorted or disappear; fail to categorize at all, and contrasts between words are hidden behind semantically
+irrelevant variation in the signal. While much attention has been paid to whether unsupervised speech models’
+(continuous or discrete) representations are predictive of phonemes, this benchmark, for the first time, explicitly
+fixes the goal of learning a discrete set of categories that are in one-to-one correspondence with the phoneme
+inventory of a language.
+
+Infants appear to learn the phoneme inventory of their language effortlessly, before they can speak. They benefit from
+millions of years of evolution of the human brain and body, giving them a learning architecture that allows them to
+thrive in the face of scarce and noisy language data, preparing them to learn the phoneme inventory of any human
+language.
+
+The Phoneme Discovery benchmark is aimed at building models that discover phoneme inventories across various languages,
+using only small amounts of speech data, and without textual data during training.
+
+## Installation
+
+```bash
+pip install discophon
+```
+
+To be able to compute ABX discriminabilities: `pip install discophon[abx]`.
+
+If you want to run baselines and have access to the utility scripts, clone this repository:
+
+```bash
+git clone https://github.com/bootphon/phoneme_discovery
+cd phoneme_discovery
+uv sync
+# uv sync --all-extras --all-groups # If you want all dependencies
+```
+
+## Usage
+
+Check out the documentation:
+
+- [Data preparation](https://github.com/bootphon/phoneme_discovery/blob/main/docs/prepare.md)
+- [Simple evaluation](https://github.com/bootphon/phoneme_discovery/blob/main/docs/evaluate.md)
+- [Run the benchmark](https://github.com/bootphon/phoneme_discovery/blob/main/benchmark.md)
+- [Use the baseline systems](https://github.com/bootphon/phoneme_discovery/blob/main/baselines.md)
+
+### Citation
+
+```bibtex
+
+```
+
+Contact: `benchmarks [at] cognitive-ml [dot] fr`

discophon-0.0.3/README.md

@@ -0,0 +1,60 @@
+# The Phoneme Discovery benchmark
+
+[💾 [Website](https://benchmarks.cognitive-ml.fr/phoneme_discovery)] [📜 [Paper]()] [📖 [BibTex](https://github.com/bootphon/phoneme_discovery?tab=readme-ov-file#citation)]
+
+## Introduction
+
+The last several years have seen revolutionary improvements in both speech processing and textual natural language
+processing. In both cases, unsupervised or self-supervised pre-training has been the key to models autonomously
+discovering representations that are tremendously useful for doing language tasks. Yet, central to the study of human
+speech processing is the phoneme inventory, a small set of discrete units that abstract away from massive pronunciation
+variability in the signal.
+
+Discovering the correct set of phonemes for a language is crucial: encode the wrong categories, and contrasts between
+words are distorted or disappear; fail to categorize at all, and contrasts between words are hidden behind semantically
+irrelevant variation in the signal. While much attention has been paid to whether unsupervised speech models’
+(continuous or discrete) representations are predictive of phonemes, this benchmark, for the first time, explicitly
+fixes the goal of learning a discrete set of categories that are in one-to-one correspondence with the phoneme
+inventory of a language.
+
+Infants appear to learn the phoneme inventory of their language effortlessly, before they can speak. They benefit from
+millions of years of evolution of the human brain and body, giving them a learning architecture that allows them to
+thrive in the face of scarce and noisy language data, preparing them to learn the phoneme inventory of any human
+language.
+
+The Phoneme Discovery benchmark is aimed at building models that discover phoneme inventories across various languages,
+using only small amounts of speech data, and without textual data during training.
+
+## Installation
+
+```bash
+pip install discophon
+```
+
+To be able to compute ABX discriminabilities: `pip install discophon[abx]`.
+
+If you want to run baselines and have access to the utility scripts, clone this repository:
+
+```bash
+git clone https://github.com/bootphon/phoneme_discovery
+cd phoneme_discovery
+uv sync
+# uv sync --all-extras --all-groups # If you want all dependencies
+```
+
+## Usage
+
+Check out the documentation:
+
+- [Data preparation](https://github.com/bootphon/phoneme_discovery/blob/main/docs/prepare.md)
+- [Simple evaluation](https://github.com/bootphon/phoneme_discovery/blob/main/docs/evaluate.md)
+- [Run the benchmark](https://github.com/bootphon/phoneme_discovery/blob/main/benchmark.md)
+- [Use the baseline systems](https://github.com/bootphon/phoneme_discovery/blob/main/baselines.md)
+
+### Citation
+
+```bibtex
+
+```
+
+Contact: `benchmarks [at] cognitive-ml [dot] fr`

discophon-0.0.3/TODO

@@ -0,0 +1,4 @@
+TODO:
+- Replace g -> ɡ in Wolof only
+- Normalize ç
+- Replace r -> ɹ in English only

discophon-0.0.3/docs/evaluate.md

@@ -0,0 +1,77 @@
+# Evaluation
+
+## Phoneme discovery
+
+You can use the `phoneme_discovery` function with `units: dict[str, list[int]]`, and `phones: dict[str, list[str]]`.
+You also need to set the number of units `n_units`, of phonemes `n_phones`, and the step (in ms) between consecutive
+units `step_units`.
+
+Example:
+
+```python
+from discophon.core import read_gold_annotations, read_submitted_units
+from discophon.evaluate import phoneme_discovery
+
+phones = read_gold_annotations("/path/to/alignments/dataset.align")
+units = read_submitted_units("/path/to/predictions/units.jsonl")
+result = phoneme_discovery(units, phones, n_units=256, n_phones=40, step_units=20)
+print(result)
+```
+
+Or via the CLI:
+
+```console
+❯ python -m discophon.evaluate --help
+usage: discophon.evaluate [-h] [--n-units N_UNITS] [--n-phones N_PHONES] [--step-units STEP_UNITS] units phones
+
+Evaluate predicted units on phoneme discovery
+
+positional arguments:
+  units                 path to predicted units
+  phones                path to gold alignments
+
+options:
+  -h, --help            show this help message and exit
+  --n-units N_UNITS     number of units
+  --n-phones N_PHONES   number of phonemes
+  --step-units STEP_UNITS
+                        step between units (in ms)
+```
+
+## ABX
+
+The ABX evaluation is done separately. First, install this package with the `abx` optional dependencies:
+
+```bash
+pip install discophon[abx]
+```
+
+Then, either run it in Python:
+
+```python
+from discophon.evaluate.abx import discrete_abx, continuous_abx
+
+result_discrete = discrete_abx("/path/to/item/dataset.item", "/path/to/predictions/units.jsonl", frequency=50)
+print("Discrete: ", result_discrete)
+
+result_continuous = continuous_abx("/path/to/item/dataset.item", "/path/to/features", frequency=50)
+print("Continuous: ", result_discrete)
+```
+
+Or via the CLI:
+
+```console
+❯ python -m discophon.evaluate.abx --help
+usage: discophon.evaluate.abx [-h] --frequency FREQUENCY item root
+
+Continuous or discrete ABX
+
+positional arguments:
+  item                  Path to the item file
+  root                  Path to the JSONL with units or directory with continuous features
+
+options:
+  -h, --help            show this help message and exit
+  --frequency FREQUENCY
+                        Units frequency in Hz
+```

discophon-0.0.3/docs/prepare.md

@@ -0,0 +1,87 @@
+# Data preparation
+
+You need the `sox` binary available in your `$PATH` for pre-processing audio files.
+
+Let's say you want to install the benchmark data and assets in a directory `$DATA`.
+
+## Download Common Voice data
+
+You first need to download audio data from CommonVoice. You can use their API if you don't want to download large
+files on your local computer.
+
+Download everything in `$DATA/raw`.
+
+Dev languages:
+
+- [Common Voice Scripted Speech 23.0 - German](https://datacollective.mozillafoundation.org/datasets/cmflnuzw5p0q7ydlq4k8skhqi) (34.41 GB)
+- [Common Voice Scripted Speech 23.0 - Swahili](https://datacollective.mozillafoundation.org/datasets/cmflnuzw7mjcay14kmowc4y96) (21.23 GB)
+- [Common Voice Scripted Speech 23.0 - Tamil](https://datacollective.mozillafoundation.org/datasets/cmflnuzw73r9g1avrbu6bwkfx) (8.56 GB)
+- [Common Voice Scripted Speech 23.0 - Thai](https://datacollective.mozillafoundation.org/datasets/cmflnuzw7fwn4fc969r5owufz) (8.35 GB)
+- [Common Voice Scripted Speech 23.0 - Turkish](https://datacollective.mozillafoundation.org/datasets/cmflnuzw71qkz8x3kil3tgjvk) (2.73 GB)
+- [Common Voice Scripted Speech 23.0 - Ukrainian](https://datacollective.mozillafoundation.org/datasets/cmflnuzw7ijdv5oe9u7ky0zrc) (2.55 GB)
+
+Test languages:
+
+- [Common Voice Scripted Speech 23.0 - Basque](https://datacollective.mozillafoundation.org/datasets/cmflnuzw5qoauo49kpf8y1gzp) (14.58 GB)
+- [Common Voice Scripted Speech 23.0 - Chinese (China)](https://datacollective.mozillafoundation.org/datasets/cmflnuzw8fvgv2vdgt6f52qvh) (21.26 GB)
+- [Common Voice Scripted Speech 23.0 - English](https://datacollective.mozillafoundation.org/datasets/cmflnuzw52mzok78yz6woemc1) (86.83 GB)
+- [Common Voice Scripted Speech 23.0 - French](https://datacollective.mozillafoundation.org/datasets/cmflnuzw5ahjms0zbrcl0vg4e) (27.87 GB)
+- [Common Voice Scripted Speech 23.0 - Japanese](https://datacollective.mozillafoundation.org/datasets/cmflnuzw5lv4n3cd25tbavjb9) (11.80 GB)
+- Wolof data comes from a different source, and will be downloaded with the other assets in the following section.
+
+Extract each archive, with `tar --strip-components=1 -xvf ...`.
+
+For example, let's say your archive is named `mcv-scripted-uk-v23.0.tar.gz`.
+Extract it with `tar --strip-components=1 -xvf mcv-scripted-uk-v23.0.tar.gz`, and move the output directory to
+`$DATA/raw`.
+
+You can delete the archives afterwards. You should have the following structure:
+
+```bash
+❯ tree -L 2 $DATA
+$DATA
+└── raw
+    ├── de
+    ├── en
+    ├── eu
+    ├── fr
+    ├── ja
+    ├── sw
+    ├── ta
+    ├── th
+    ├── tr
+    ├── uk
+    └── zh-CN
+```
+
+## Download benchmark assets
+
+Now download the benchmark assets with the following command:
+
+```bash
+python -m discophon.prepare download $DATA
+```
+
+This will download:
+- Symlinks to audio files for each split in each language
+- Manifests
+- Alignments and item files
+
+## Preprocess selected audio files
+
+Now resample audio files and convert them to WAV with the command:
+
+```bash
+for code in swa tam tha tur ukr cmn eus jpn; do
+    python -m discophon.prepare audio $DATA $code
+done
+```
+
+This will create directories `$DATA/audio/cmn/all`, `$DATA/audio/deu/all`, `$DATA/audio/eng/all`, etc. with
+resampled audio files. The directories corresponding to each split contain symlinks to those files.
+
+You should parallelize this loop if you can. If you are in a SLURM cluster, you should also parallelize each dataset
+processing across tasks or array jobs. The `discophon.prepare` package will automatically handle the distribution of
+files to process across jobs.
+
+You can delete the `$DATA/raw` folder afterwards.

discophon-0.0.3/pyproject.toml

@@ -0,0 +1,116 @@
+[project]
+name = "discophon"
+description = "The Phoneme Discovery Benchmark"
+readme = "README.md"
+requires-python = ">=3.12"
+license = "MIT"
+authors = [{ name = "CoML", email = "dev@cognitive-ml.fr" }]
+dependencies = [
+  "filelock>=3.20.2",
+  "httpx>=0.28.1",
+  "joblib>=1.5.3",
+  "numba>=0.63.1",
+  "numpy>=2.3.5",
+  "polars>=1.36.1",
+  "praat-textgrids>=1.4.0",
+  "soundfile>=0.13.1",
+  "soxr>=1.0.0",
+  "tqdm>=4.67.1",
+  "xarray>=2025.12.0",
+]
+dynamic = ["version"]
+
+[project.optional-dependencies]
+abx = [
+  "fastabx>=0.6.1",
+]
+
+[dependency-groups]
+dev = [
+  "ipykernel>=7.1.0",
+  "ruff>=0.14.10",
+  "tombi>=0.7.16",
+  "ty>=0.0.11",
+  "typos>=1.42.0",
+]
+baselines = [
+  "minimal-hubert>=0.0.1",
+  "spidr[train]>=0.1.3",
+]
+plot = [
+  "altair>=6.0.0",
+  "matplotlib>=3.10.8",
+  "seaborn>=0.13.2",
+  "vegafusion>=2.0.3",
+  "vl-convert-python>=1.9.0",
+]
+test = [
+  "matplotlib>=3.10.8",
+  "pytest>=9.0.2",
+]
+
+[build-system]
+requires = ["hatch-vcs", "hatchling"]
+build-backend = "hatchling.build"
+
+[tool.hatch.version]
+source = "vcs"
+
+[tool.pytest]
+log_cli = true
+log_cli_level = "INFO"
+
+[tool.ruff]
+line-length = 119
+
+[tool.ruff.lint]
+select = [
+  "A",  # flake8-builtins
+  "ARG",  # flake8-unused-arguments
+  "ANN",  # flake8-annotations
+  "B",  # flake8-bugbear
+  "BLE",  # flake8-blind-except
+  "C4",  # flake8-comprehensions
+  "C90",  # mccabe
+  "E",  # pycodestyle error
+  "EXE",  # flake8-executable
+  "F",  # Pyflakes
+  "FBT",  # flake8-boolean-trap
+  "FLY",  # flynt
+  "FURB",  # refurb
+  "G",  # flake8-logging-format
+  "I",  # isort
+  "ICN",  # flake8-import-conventions
+  "ISC",  # flake8-implicit-str-concat
+  "LOG",  # flake8-logging
+  "N",  # pep8-naming
+  "NPY",  # NumPy-specific rules
+  "PERF",  # Perflint
+  "PIE",  # flake8-pie
+  "PGH",  # pygrep-hooks
+  "PL",  # Pylint
+  "PTH",  # flake8-use-pathlib
+  "Q",  # flake8-quotes
+  "RET",  # flake8-return
+  "RSE",  # flake8-raise
+  "RUF",  # Ruff-specific rules
+  "SIM",  # flake8-simplify
+  "SLF",  # flake8-slots
+  "T10",  # flake8-debugger
+  "TC",  # flake8-type-checking
+  "TD",  # flake8-todos
+  "TID",  # flake8-tidy-imports
+  "UP",  # pyupgrade
+  "W",  # pycodestyle warning
+]
+
+[tool.ruff.lint.flake8-tidy-imports]
+banned-module-level-imports = ["discophon.evaluate.abx"]
+
+[tool.ruff.lint.pylint]
+allow-magic-value-types = ["int", "str"]
+max-args = 10
+max-statements = 75
+
+[tool.typos.default.extend-words]
+tha = "tha"

discophon-0.0.3/scripts/benchmark.slurm

@@ -0,0 +1,65 @@
+#!/bin/bash
+#SBATCH -J discophon
+#SBATCH --cpus-per-task=10
+#SBATCH --time=05:00:00
+#SBATCH --array=1-12
+
+set -euo pipefail
+
+usage() {
+    cat << EOF
+Usage: $0 DATASET UNITS FEATURES OUTPUT N_UNITS [STEP_UNITS]
+
+Arguments:
+    DATASET      Path to the dataset
+    UNITS        Path to the discrete units (used for benchmarks 'discovery' and 'abx-discrete')
+    FEATURES     Path to the continuous features (used for benchmark 'abx-continuous')
+    OUTPUT       Output JSONL file
+    N_UNITS      Number of units
+    STEP_UNITS   Step size for units (default: 20)
+
+Description:
+    Runs all tasks in the Phoneme Discovery benchmark, across all languages and splits combinations.
+    Array tasks 1-4:  phoneme discovery
+    Array tasks 5-8:  ABX discrete
+    Array tasks 9-12: ABX continuous
+
+Example:
+    sbatch $0 ./path/to/dataset ./path/to/units ./path/to/features ./path/to/output 256
+EOF
+}
+
+if [[ "${1:-}" == "-h" || "${1:-}" == "--help" ]]; then
+    usage
+    exit 0
+fi
+
+if [[ $# -lt 5 ]]; then
+    echo "Error: Missing required arguments." >&2
+    usage
+    exit 1
+fi
+
+DATASET=$1
+UNITS=$2
+FEATURES=$3
+OUTPUT=$4
+N_UNITS=$5
+STEP_UNITS=${6:-20}
+
+benchmark=$(( (SLURM_ARRAY_TASK_ID - 1) / 4 ))
+benchmark=$(echo "discovery abx-discrete abx-continuous" | cut -d' ' -f$((benchmark + 1)))
+[[ "$benchmark" == "abx-continuous" ]] && predictions="$FEATURES" || predictions="$UNITS"
+task=$(( (SLURM_ARRAY_TASK_ID - 1) % 4 ))
+[[ $task -lt 2 ]] && split="dev" || split="test"
+[[ $((task % 2)) -eq 0 ]] && languages="dev" || languages="test"
+
+srun python -m discophon.benchmark \
+  "$DATASET" \
+  "$predictions" \
+  "$OUTPUT" \
+  --languages "$languages" \
+  --split "$split" \
+  --benchmark "$benchmark" \
+  --n-units "$N_UNITS" \
+  --step-units "$STEP_UNITS"