Error code:   ConfigNamesError
Exception:    ValueError
Message:      
Expected data_files in YAML to be either a string or a list of strings
or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'DarijaCS-Prism', 'path': 'data/DarijaCS-Prism/*.parquet'}, {'split': 'DarijaCS-Probe', 'path': 'data/DarijaCS-Probe/*.parquet'}]
Examples of data_files in YAML:

   data_files: data.csv

   data_files: data/*.png

   data_files:
    - part0/*
    - part1/*

   data_files:
    - split: train
      path: train/*
    - split: test
      path: test/*

   data_files:
    - split: train
      path:
      - train/part1/*
      - train/part2/*
    - split: test
      path: test/*

PS: some symbols like dashes '-' are not allowed in split names

Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/dataset/config_names.py", line 67, in compute_config_names_response
                  config_names = get_dataset_config_names(
                      path=dataset,
                      token=hf_token,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 161, in get_dataset_config_names
                  dataset_module = dataset_module_factory(
                      path,
                  ...<4 lines>...
                      **download_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1182, in dataset_module_factory
                  raise e1 from None
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 1157, in dataset_module_factory
                  ).get_module()
                    ~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/load.py", line 613, in get_module
                  metadata_configs = MetadataConfigs.from_dataset_card_data(dataset_card_data)
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/metadata.py", line 153, in from_dataset_card_data
                  cls._raise_if_data_files_field_not_valid(metadata_config)
                  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/utils/metadata.py", line 100, in _raise_if_data_files_field_not_valid
                  raise ValueError(yaml_error_message)
              ValueError: 
              Expected data_files in YAML to be either a string or a list of strings
              or a list of dicts with two keys: 'split' and 'path', but got [{'split': 'DarijaCS-Prism', 'path': 'data/DarijaCS-Prism/*.parquet'}, {'split': 'DarijaCS-Probe', 'path': 'data/DarijaCS-Probe/*.parquet'}]
              Examples of data_files in YAML:
              
                 data_files: data.csv
              
                 data_files: data/*.png
              
                 data_files:
                  - part0/*
                  - part1/*
              
                 data_files:
                  - split: train
                    path: train/*
                  - split: test
                    path: test/*
              
                 data_files:
                  - split: train
                    path:
                    - train/part1/*
                    - train/part2/*
                  - split: test
                    path: test/*
              
              PS: some symbols like dashes '-' are not allowed in split names

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

DarijaCS

Two complementary benchmarks for Moroccan Darija code-switched automatic speech recognition

DarijaCS-Prism · Broad evaluation across domains and acoustic conditions
DarijaCS-Probe · Targeted evaluation to discover model weaknesses

Developed and released by the Ministry of Artificial Intelligence, Digital Transformation and Administration Reform, Kingdom of Morocco.

Dataset: https://huggingface.co/datasets/Morocco-MITNRA-Labs/DarijaCS-asr-eval
License: Creative Commons Attribution 4.0 International (CC BY 4.0) — free to use, including commercially, with attribution.

Overview

Moroccan Darija is dynamic, multilingual, and highly context-dependent. Speakers frequently move between Darija and other languages or language varieties, sometimes within a single sentence. Speech recognition systems must also contend with differences in vocabulary, subject matter, speaking rate, speaker characteristics, recording quality, and background noise.

Yet automatic speech recognition (ASR) is often evaluated through a single aggregate score. Such a score can obscure substantial variation in performance and provide little insight into why a system fails.

DarijaCS is an open benchmark initiative developed by our internal team at the Ministry to make the evaluation of Moroccan Darija ASR more comprehensive, transparent, and useful to the wider ecosystem. It brings together two complementary evaluation perspectives:

DarijaCS-Prism: How does an ASR system perform across a broad and diverse range of Moroccan speech?
DarijaCS-Probe: Which specific phenomena expose weaknesses or recurring failure patterns in an ASR system?

CS stands for code switching, the alternation between languages or language varieties in speech.

Our objective is to provide researchers, open-source contributors, startups, industry teams, and public institutions with shared evaluation resources that support better speech technologies for Morocco.

Access the dataset

Both benchmarks are available in the official DarijaCS Hugging Face repository.

The benchmarks

DarijaCS-Prism

A Multi-Domain, Multi-Condition Benchmark for Moroccan Darija Code-Switched ASR

A prism reveals multiple facets of the same signal. Likewise, DarijaCS-Prism is designed to reveal different facets of speech recognition performance across a deliberately broad and diverse set of Moroccan speech samples.

Its coverage includes variation along several dimensions:

Dimension	Evaluation focus
Domains	A broad variety of subjects, contexts, and spoken content
Code switching	Different language-mixing patterns and transitions within speech
Speech rates	Different speaking speeds and delivery styles
Acoustic conditions	Variation in noise, recording conditions, and audio quality
Speakers	Diversity in voices and speech characteristics

Prism is intended to help assess breadth and robustness. An overall ASR score is useful, but performance differences across domains, speakers, code-switching patterns, and acoustic conditions are often equally important.

What Prism tells you: how a model behaves across a varied collection of speech conditions, and where its performance changes across those conditions.

DarijaCS-Probe

A Targeted Diagnostic Benchmark for Discovering Weaknesses in Moroccan Darija Code-Switched ASR

DarijaCS-Probe takes a different approach. Rather than aiming primarily for broad coverage, it focuses on targeted evaluation cases selected to reveal potential blind spots in ASR systems.

Probe is useful for investigating particular recognition behaviors, analyzing recurring errors, and testing whether model improvements address identified weaknesses. It is designed to support diagnosis rather than compress every aspect of performance into a single ranking.

What Probe tells you: which targeted patterns warrant closer investigation, and whether a model handles those patterns consistently.

Important distinction: Probe is not presented as an inherently harder benchmark. Its samples are selected for diagnostic value, not because they necessarily belong to a more difficult or statistically representative distribution. A higher error rate on Probe does not, by itself, establish that Probe is universally harder than Prism.

Why both matter

Evaluation question	Benchmark
How well does a model handle varied Moroccan Darija speech?	Prism
How does performance vary across domains, speaking styles, or recording conditions?	Prism
Where are a model's specific blind spots or recurring mistakes?	Probe
Did a model update improve broad coverage and address targeted weaknesses?	Prism + Probe, reported separately

The two benchmarks are complementary, not interchangeable. We recommend reporting them independently rather than merging them into one headline score.

Intended use

DarijaCS is intended for:

Benchmarking: comparing Moroccan Darija ASR systems under a common evaluation approach.
Model development: identifying areas where additional research or engineering may improve speech recognition.
Error analysis: understanding the effects of code switching, speaking patterns, and acoustic conditions on recognition accuracy.
Research: studying multilingual and code-switched speech recognition in the Moroccan context.
Community progress: enabling reproducible experiments and more informative reporting across the ecosystem.

Both academic and commercial use are welcome under the license below.

Evaluation recommendations

The benchmark task is automatic speech recognition: given an audio recording, generate a transcription and compare it with the reference transcription.

We recommend reporting Word Error Rate (WER) and Character Error Rate (CER) for each benchmark separately.

WER measures word-level substitutions, deletions, and insertions relative to the reference transcription.
CER measures character-level substitutions, deletions, and insertions; it can provide additional insight where orthographic conventions vary.

For edit-distance metrics, the error rate is calculated as:

$\mathrm{Error\ Rate} = \frac{S + D + I}{N}$

where S is substitutions, D is deletions, I is insertions, and N is the number of reference words (for WER) or characters (for CER).

Recommended reporting practices

Report Prism and Probe separately. Avoid combining their scores into an overall DarijaCS score.
Use consistent preprocessing. Apply the same normalization rules to predictions and references for all systems being compared.
Document transcription conventions. In particular, explain how Arabic and Latin scripts, punctuation, numbers, and alternative spellings are handled.
Include breakdowns where supported by the released annotations. Aggregate scores may hide variation by domain, acoustic condition, speech rate, or diagnostic phenomenon.
Document model settings. Report the model name and version, decoding strategy, preprocessing, normalization, and any task-specific adaptation.
Disclose benchmark exposure. Make clear whether evaluation audio or reference transcripts were used during training, tuning, or model selection.

Interpreting results

Lower WER and CER are better, provided the same scoring protocol is used.
Results produced with different normalization procedures may not be directly comparable.
Prism's variety supports broad evaluation, but diversity is not the same as statistical representativeness of every Moroccan speech setting.
Probe scores measure performance on the selected diagnostic cases; they should not be treated as estimates of average real-world error rates.
Neither benchmark alone replaces evaluation on audio collected from a system's actual deployment environment.

Reproducibility and evaluation integrity

To support meaningful comparisons:

Run the ASR model on the benchmark audio without providing the reference transcription as model input.
Align predictions with the corresponding reference transcriptions.
Apply a clearly documented and consistent normalization procedure.
Calculate WER and CER independently for DarijaCS-Prism and DarijaCS-Probe.
Report the results alongside the evaluation configuration and any relevant disaggregated analysis.

Do not describe evaluation results as independent held-out performance if the same references have been used to train, tune, or select the model. Transparent disclosure is essential for fair comparison.

Limitations and responsible use

Moroccan Darija is linguistically rich, and no fixed benchmark can capture every speaker, region, accent, setting, or code-switching phenomenon. Transcription conventions can also affect measured accuracy, especially when speech can reasonably be written in multiple ways.

The benchmarks are tools for evaluation and diagnosis—not exhaustive descriptions of all Moroccan speech or guarantees of deployment performance.

We encourage users to handle speech recordings responsibly, respect applicable privacy and data-protection laws, and avoid attempts to identify speakers or infer sensitive personal attributes from recordings. These are responsible-use recommendations, not additional copyright-license restrictions.

Open license

DarijaCS-Prism and DarijaCS-Probe are released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

This means anyone can use the benchmarks, including for research, education, commercial products, model evaluation, training, modification, and redistribution, without requesting permission or paying licensing fees, provided the license conditions are respected.

The main requirement is attribution:

Credit the Ministry of Artificial Intelligence, Digital Transformation and Administration Reform, Kingdom of Morocco as the source.
Identify the benchmark used: DarijaCS-Prism, DarijaCS-Probe, or both.
Include a link to the CC BY 4.0 license and indicate whether changes were made, as required by that license.

Attribution does not imply endorsement by the Ministry. The license applies to the rights held by the releasing institution and does not override other applicable rights, such as privacy or personality rights.

Project Leadership and Acknowledgments

DarijaCS-Prism and DarijaCS-Probe were developed by the internal AI team of the Ministry of Artificial Intelligence, Digital Transformation and Administration Reform of Morocco.

The initiative was conceived, coordinated, and supervised by Abderrahman Skiredj, AI Expert and Advisor to the Minister.

We acknowledge the contributions of the Ministry's AI team in dataset development, curation, and evaluation.

How to cite

If you use DarijaCS in a publication, product, report, presentation, model card, or other work, please cite the Ministry and the benchmark name.

Suggested citation:

Ministry of Artificial Intelligence, Digital Transformation and Administration Reform, Kingdom of Morocco (2026). DarijaCS: DarijaCS-Prism and DarijaCS-Probe — Benchmarks for Moroccan Darija Code-Switched Automatic Speech Recognition. Available at: https://huggingface.co/datasets/Morocco-MITNRA-Labs/DarijaCS-asr-eval

BibTeX:

@misc{darijacs2026,
  title        = {DarijaCS: DarijaCS-Prism and DarijaCS-Probe --- Benchmarks for Moroccan Darija Code-Switched Automatic Speech Recognition},
  author       = {{Ministry of Artificial Intelligence, Digital Transformation and Administration Reform, Kingdom of Morocco}},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/Morocco-MITNRA-Labs/DarijaCS-asr-eval},
  note         = {DarijaCS-Prism: multi-domain, multi-condition evaluation; DarijaCS-Probe: targeted weakness discovery. Licensed under CC BY 4.0}
}

Community

DarijaCS is an initiative built by the Ministry's internal team and shared openly to support a stronger Moroccan AI ecosystem.

We welcome the community's contributions through reproducible benchmark results, transparent error analyses, feedback on evaluation practices, and research that helps improve Moroccan Darija speech recognition. We particularly encourage comparisons that explain where models improve or regress, rather than focusing exclusively on headline scores.

Our ambition is simple: make progress in Moroccan speech AI easier to measure, understand, and share.

DarijaCS-Prism reveals the breadth. DarijaCS-Probe reveals the blind spots.

Developed in Morocco. Open to everyone.

Downloads last month: -

Total file size:

173 MB