sawalni-ai/fw-darija

Gherbal’ing Multilingual Fineweb 2 🍵

Following up on their previous release, the fineweb team has been hard at work on their upcoming multilingual fineweb dataset which contains a massive collection of 50M+ sentences across 100+ languages. The data, sourced from the Common Crawl corpus, has been classified into these languages using GlotLID, a model able to recognize more than 2000 languages. The performance of GlotLID is quite impressive, considering the complexity of the task. It is able to correctly identify the language of a sentence with a decent degree of accuracy. However, especially for low-resource languages, it still makes some mistakes, and some languages are more difficult for it to identify than others.

This caught our interest and we wanted to see if we could help improve the quality of the dataset using our Gherbal language identification model, which we recently released and made available on our API platform. Our performance on several low-resource languages, notably Moroccan, Persian and Swahili, is quite impressive, and the hope is to expand the available resources for these underserved communities.

We gladly took on the challenge and as a first step, we chose to focus on Moroccan Arabic, a language spoken by millions of people in Morocco and people of Moroccan descent in Europe. This report will detail the process we went through to achieve our goal, and the results we obtained.

The Dataset

The dataset we were given is a bunch of parquet files containing 50M+ sentences, with the following columns:

• id: a unique identifier for the document
• text: the document itself, extracted from a webpage
• metadata: a json column containing metadata about the sentence, including the url of the page it was found on, the date of the page, and the previous classification of the page by GlotLID. The dataset contains several configurations, each one corresponding to a different language. We focused our attention on the Arabic arb_Arab_dedup and the Moroccan ary_Arab_dedup configurations, which we will refer to as the “Arabic” and “Moroccan” datasets respectively.

Our Approach

Dataset Processing

To tackle this challenge, we developed a systematic pipeline to process and analyze each dataset. First, we performed thorough text cleanup to remove any unwanted artifacts and standardize the content. This ensures we’re working with clean, natural language text, especially as webpage content can be quite noisy. Next, we leveraged the Natural Language Toolkit (NLTK) library to break down the documents into individual sentences. While this is not a perfect solution and the noisy content can make it difficult to identify sentences particulary for languages not supported by NLTK, it is a good enough approximation for our purposes, which is to reduce the variance in a webpage and avoid confusing the model with extremely long mixed-language content. This step was crucial as it allowed us to analyze the text at a more granular level. With our sentences prepared, we then ran each one through our Gherbal language detection model. The model evaluated each sentence and provided a confidence score across the 33 languages Gherbal supports. We then aggregated these sentence-level results by averaging the classification scores. This gave us a comprehensive understanding of the dominant language patterns within each document. A more fine-grained analysis at the sentence level would have yielded more data with higher quality, but ultimately postponed to a later release given the time and resource constraints. Finally, we applied a filtering step to focus specifically on content classified as Moroccan Arabic in Arabic script (ary_Arab). The resulting dataset is available on Huggingface at sawalni-ai/ml-fw-darija.

Dataset Analysis

We used our Klimat library to analyze the dataset. Klimat is a tool we developed to perform statistical analysis on language datasets, and is able to generate a number of interesting insights into the data. We will share more about Klimat in a future blog post, but for now we will focus on the results we obtained for the fineweb dataset.

Website Analysis

We also performed an analysis on the websites that were used to source the data in multilingual fineweb, and classified by Gherbal as Moroccan Arabic. This gave us an interesting insight on where Moroccan Arabic is used on the web, which could be useful to increase the quantity of high quality data for the language. We broke down the data by multiple criteria, including the top level domain, the duration the website was online (based on Common Crawl accessing it), and more. We restricted the analysis to high confidence samples, and filtered to the top 1000 websites by quantity of data.

Our Results

Let’s start by looking at the results for the Moroccan dataset.

• Original count in ary_Arab: 5.8M
• Resulting count after filtering: 37352 (0.64% of original)
• Number of tokens in ary_Arab: 2.8B (estimated using tiktoken for multilinguality)
• Number of tokens in filtered dataset: 75.3M

False Positives

A manual review of the filtered dataset revealed that human preferences were consistent with the results of Gherbal, and that the filtered dataset should be a good resource for training and evaluating models for Moroccan, despite the small sample size. It is worth noting that Algerian and Tunisian Arabic were also misclassified as Moroccan Arabic due to the elevated mutual intelligibility between the three. This is a known current limitation of Gherbal which only supports Moroccan and Egyptian varieties of Arabic and should be addressed in future releases.

False Negatives

Looking at our Gherbal paper (pending publication), specifically at the benchmark results on the flores-200 devtest set, we can estimate that the false negative rate from Standard Arabic (arb_Arab) to Moroccan Arabic (ary_Arab) is around 10%. Extrapolating this figure, we can estimate the false negative rate for the filtered dataset to be around 37352 * 0.1 = 3735 Moroccan Arabic sentences that were incorrectly filtered out.

Other dataset configurations

We also applied the same process to the other dataset configurations, namely the Arabic (arb_Arab) and the latin-script Arabic (arb_Latn) configurations. While the results are not yet complete, we can already draw some interesting observations:

• Arabic (arb_Arab)
  • Original count in arb_Arab: 24.2M
  • Resulting count after filtering: 0 While some samples (<100 in total) were classified as Moroccan Arabic, a manual review revealed that these were all incorrect classifications by Gherbal and that the filtered dataset is indeed empty. This might change as we process the rest of the dataset, or as we improve Gherbal’s performance on Arabic and its related languages. The resulting dataset will be made available as an additional configuration on the same dataset here when the processing is complete.
• Arabic in Latin script (arb_Latn)
  • Original count in arb_Latn: 600K
  • Resulting count after filtering: 15K (2.5% of original) This dataset is classified as arb_Latn by GlotLID, which presents extreme variance in the data as Arabic can be transliterated in so many different ways. As Gherbal is able to correctly identify ary_Latn (Moroccan Arabic in Latin script) with a high degree of accuracy, we are able to recover a significant amount of data that was previously quite unusable among all the noise. We also observe that this dataset contains the most variations in the actual language as classified by Gherbal, which confirms that the arb_Latn label from GlotLID is not a good proxy for high quality Arabic data in latin script. The resulting dataset will be made available as an additional configuration on the same dataset here when the analysis is complete.

Team

This project was conducted by Omneity Labs:

• Omar Kamali
• Abdeljalil Elmajjodi
• Mohamed Abchir

Omneity Labs (Sawalni team) is a Moroccan private R&D lab specialized in low-resource languages and cultural alignment. We build AI tools and products for low-resource languages and underserved communities.