WolBanking77: Wolof Intent Dataset

• WolBanking77 is a low-resource dataset featuring 9,791 bilingual Wolof-French text examples and over 4 hours of speech recordings for banking intent detection.
• The resource covers 77 text intents and 10 speech intents, supporting comprehensive evaluations in intent classification, ASR, and cross-modal analysis.
• Data collection involved expert translations, localized recordings in Senegal, and baseline evaluations using both classical and transformer-based models.

Searching arXiv for the two referenced papers to ground the article in current sources. WolBanking77 is a Wolof intent classification dataset for academic research on low-resource NLP and speech, with explicit emphasis on Senegalese Wolof and on voice-oriented access to banking and related services. It is presented as a bimodal resource: a text corpus containing 9,791 sentences in the banking domain and a speech corpus containing more than 4 hours of spoken sentences. The text portion spans 77 intent classes, while the audio portion is smaller and covers 10 intents. The dataset is framed as a response to the scarcity of Wolof benchmarks for intent detection, spoken language understanding, and ASR, particularly in a setting where Wolof is spoken by around 90\% of the population of Senegal and Senegal has an illiteracy rate of 42\% according to the World Bank (Kandji et al., 23 Sep 2025).

1. Corpus identity and research motivation

WolBanking77 is positioned as a low-resource benchmark for Senegalese Wolof, with the stated goal of supporting intent detection models in both text and speech. Its central motivation is the mismatch between the demand for conversational or voice-based systems and the limited availability of Wolof language resources. The paper links this motivation to practical access constraints: many users may interact with financial or public-service systems primarily through speech rather than text, especially in informal economic settings and under low-literacy conditions (Kandji et al., 23 Sep 2025).

The dataset is specifically aligned with intent classification. In its text modality, each instance consists of a sentence and an intent label, with Wolof and French versions. In its speech modality, each utterance is accompanied by a transcription and an intent label, enabling ASR and spoken intent-related experiments. This dual design makes the resource relevant not only for standard text classification but also for end-to-end speech pipelines and comparative cross-modal evaluation.

A useful distinction is that WolBanking77 is a banking-centered resource in its text component, whereas its speech subset is smaller and includes both e-banking and transport-related intents. This suggests that the dataset family is organized around practical spoken-assistant use cases rather than around a single narrowly delimited ontology.

2. Dataset composition, label space, and split protocol

The text component contains 9,791 customer-service queries translated from the English Banking77 training set and manually localized into French and Wolof. It spans 77 intent classes, matching the original Banking77 intent inventory. The paper reports two text configurations: a 5k sample and the full 9,791-example version. For both, the split is 80/20 train/test, and no separate development set is described in the main text. The 5k version is split into 4,000 train and 1,000 test examples; the full version is split into 7,832 train and 1,959 test examples (Kandji et al., 23 Sep 2025).

The label distribution is imbalanced. The most frequent intent appears 200 times, and the least frequent one 24 times. The paper also reports sentence-length statistics. For Wolof, sentence length ranges from 2 to 81 words, with mean 12.22 and median 10. For French, the range is 2 to 83 words, with mean 12.47 and median 10. Duplicate Wolof translations were removed for intent-detection experiments in order to avoid many-to-one translation collisions from English into Wolof.

The speech component is smaller but technically specific. It contains 263 utterances covering 10 intents, for a total duration of about 4 hours and 17 minutes. The recordings are WAV, single-channel, sampled at 16 kHz, with an average utterance duration of 4,815 ms. The audio set was derived from MINDS-14 and extended with additional intents. The paper lists OPEN_ACCOUNT, BUS_RESERVATION, TECHNICAL_VISIT, TRANSFER_MONEY, and AMOUNT as added intents, alongside BALANCE, CASH_DEPOSIT, FREEZE, LATEST_TRANSACTIONS, and PAY_BILL (Kandji et al., 23 Sep 2025).

Modality	Scale	Notes
Text	9,791 sentences	77 intents; Wolof and French
Speech	263 utterances	10 intents; about 4 h 17 min

This composition is significant because the text corpus is sufficiently large to support broad intent-classification benchmarking, while the audio subset provides a compact but operational testbed for ASR and spoken intent pipelines.

3. Translation, localization, and recording workflow

Data creation was explicitly human-centered. For the text portion, customer-service queries from Banking77 were manually translated into French and Wolof by linguistic experts from the Centre de Linguistique Appliquée de Dakar (CLAD). The Wolof translations were localized to the Senegalese context; the paper gives examples such as “ATM” becoming “GAB” and “app” becoming “aplikaasiyoŋ” (Kandji et al., 23 Sep 2025).

For the speech portion, the text queries were translated and transcribed, and phonetic transcription in Wolof was provided to help speakers pronounce the sentences correctly. Recordings were collected from students at Cheikh Anta Diop University (UCAD) in Dakar, in a closed environment, using the Lig-Aikuma Android app in elicitation mode on a tablet. Before recording, each participant filled in a metadata form containing information such as native language, region of origin, gender, and year of birth. A user_id was generated to anonymize the speaker name, and the workflow allowed a sentence to be cancelled if it was mispronounced.

In total, 31 recording sessions were conducted, involving 14 male, 14 female, and 3 unspecified speakers. Corrupted audio files were removed. The datasheets state that the audio was recorded in an uncontrolled environment and therefore contains background noise. The paper also notes that punctuation and numbers were removed and text was lowercased before splitting, and that special characters were removed for ASR evaluation. Python was used for cleaning, but the authors do not release all raw data and metadata; they release the curated dataset.

These design choices indicate that WolBanking77 is not merely a translated benchmark but a localized and partially speech-engineered resource. A plausible implication is that its value lies as much in controlled annotation and documentation practice as in raw example count.

4. Intent-classification baselines and evaluation regime

For text intent detection, the evaluation metric is weighted average F1-score, with precision and recall also reported for some models. The benchmark suite includes classic machine-learning baselines on bag-of-words representations, LASER3 sentence embeddings with downstream classifiers, pretrained transformer models evaluated in zero-shot, few-shot, and fine-tuned settings, and prompt-based fine-tuning with Llama-3.2 models (Kandji et al., 23 Sep 2025).

The classic baselines are k-nearest neighbors, support vector machine, logistic regression, and naive Bayes. LASER3 is followed by either an MLP or a CNN classifier trained from scratch, and the CNN is also examined in a tuned configuration using Ray Tune with ASHAScheduler. The transformer set includes BERT base uncased, AfroXLMR-large, AfroLM_active_learning, mDeBERTa-v3-base-mnli-xnli, and AfriTeVa V2 Base. Few-shot experiments use SetFit, with 2-shot and 8-shot meaning 2 or 8 examples per intent. The paper also benchmarks Llama-3.2-1B-Instruct and Llama-3.2-3B-Instruct via prompt-based fine-tuning.

The reported results show a sharp separation between zero-shot transfer and supervised adaptation. Zero-shot scores are very low across the pretrained encoders. Fine-tuning materially improves performance, and the best reported text result comes from fine-tuned AfroXLMR: F1 = 79 on the 5k subset and F1 = 57\% on the full 9,791-example set. Among the classic methods, BoW + SVM and BoW + LR reach F1 = 0.68 on the 5k split, while on the full split BoW + LR reaches F1 = 0.53. Among the LASER3-based methods, LASER3 + MLP performs best, with F1 = 0.55 on the 5k split and 0.42 on the full split, whereas LASER3 + CNN performs very poorly unless tuned, and even the tuned version does not surpass the MLP.

Setting	Best reported model	Result
5k text intent detection	Fine-tuned AfroXLMR	F1 = 79
Full text intent detection	Fine-tuned AfroXLMR	F1 = 57%
Speech ASR	Canary-1b-flash	WER = 0.59%

The paper also reports prompt-based small-language-model baselines. On the 5k split, Llama-3.2-3B-Instruct obtains precision 0.76, recall 0.75, and F1 0.75; on the full split it obtains precision 0.56, recall 0.55, and F1 0.55. This places it close to the strongest fine-tuned text classifiers on the full setting, although the paper’s strongest figure-level result remains AfroXLMR.

5. Speech subset and ASR benchmarking

The ASR experiments use Word Error Rate (WER) on the approximately 4-hour audio subset. The baselines are Canary-1b-flash, Phi-4-multimodal-instruct, and Distil-whisper-large-v3.5. All three models are fine-tuned on the WolBanking77 audio dataset for 1,000 steps, with experiments run in PyTorch using Hugging Face Transformers and NVIDIA NeMo where appropriate (Kandji et al., 23 Sep 2025).

The reported WER values are 0.59\% for Canary-1b-flash, 3.1\% for Phi-4-multimodal-instruct, and 4.63\% for Distil-whisper-large-v3.5. The paper emphasizes that Canary-1b-flash is the best ASR model by a large margin and highlights that strong ASR performance is achievable even with only about 4 hours of Wolof speech. Training-time details are also given: Canary-1b-flash is reported at 20 minutes, Phi-4-multimodal-instruct at 32 minutes, and Distil-whisper-large-v3.5 at 44 minutes.

The speech benchmark is methodologically narrower than the text benchmark, but it is notable for tying intent-classification research to actual speech processing in Wolof. Because the paper also notes background noise, incomplete dialect coverage, and speaker skew toward younger participants, these WER results should be interpreted as strong baseline performance within the recorded conditions rather than as a comprehensive characterization of Wolof ASR in the wild.

6. Relation to earlier Wolof intent datasets, limitations, and release conditions

WolBanking77 should be distinguished from an earlier Wolof intent-classification resource introduced in work on task-oriented dialog systems for Senegalese Wolof (Mbaye et al., 2024). That earlier resource was not called “WolBanking77.” It was a synthetic projected dataset derived from a French subset of MASSIVE, covering 10 domains, 27 intents, and 9,638 examples, with translation-based annotation projection and a Rasa pipeline using LaBSE embeddings and DIET. Its intent-classification results were near parity between French and synthetic Wolof, with overall macro F1 of 0.999 for French and 0.995 for Wolof. By contrast, WolBanking77 is a later, separately named dataset centered on Banking77-derived text data and a distinct speech component. A common misconception is to conflate the two because both concern Wolof intent classification; the papers describe materially different resources, construction procedures, and task scopes (Mbaye et al., 2024).

The limitations of WolBanking77 are stated explicitly. They include label imbalance in the intent classes, small speech size, background noise from uncontrolled recording conditions, and limited dialect and accent coverage. The paper also notes that the text corpus is a translated version of Banking77, so some distributional quirks may remain from the source corpus. For the speaker sample, age statistics are summarized as mean 26, standard deviation 4, minimum 22, 25\% 23, median 25, 75\% 27, and maximum 36; older speakers are underrepresented, and the authors state that they intend to include more older voices in future versions (Kandji et al., 23 Sep 2025).

On privacy and release conditions, the paper reports that speaker names, location, and androidID were removed, and that consent forms were used. No confidential or offensive content is reported. Licensing is CC BY 4.0 for both dataset and source code, permitting sharing and adaptation with attribution, including commercial use. The dataset is planned for public internet release and DOI distribution, with maintenance and updates to be communicated via Kaggle. The authors further indicate that older versions will be kept and that future updates may add more audio recordings, improve the simulation scripts, include response texts for TTS use, and potentially add slot-filling annotations.

These release terms and documented limitations position WolBanking77 as a practical benchmark rather than a finalized or exhaustive standard. Its main research utility lies in combining a sizeable banking-domain Wolof text corpus with a smaller but operational speech benchmark, under transparent documentation of preprocessing, recording conditions, and baseline behavior.