KabSentiment

A 3-class sentiment benchmark for Kabyle (Taqbaylit, kab, Latin script), from the AƔBALU project.

15,000 sentences drawn from human-written Kabyle text, labelled via cross-lingual teacher projection and balanced across negative, neutral, and positive classes.

from datasets import load_dataset

ds = load_dataset("agbalu/KabSentiment")

Splits

Split	Sentences	Negative	Neutral	Positive
train	12,000	4,007	4,000	3,993
dev	1,500	472	510	518
test	1,500	521	490	489
Total	15,000	5,000	5,000	5,000

The dataset contains exactly 5,000 sentences per class across the entire 15,000-sentence collection. Splits were partitioned via corpus-wide stratified shuffling, resulting in minor class variations within individual splits (majority class baselines are 34.5% on dev and 34.7% on test).

Schema

Field	Type	Description
id	string	Unique record ID (kab_sent_{split}_{idx})
text_kab	string	Normalized Kabyle sentence
label	int	0 negative, 1 neutral, 2 positive
label_name	string	Label string (negative, neutral, positive)
confidence_score	float	Classifier prediction probability for assigned class (≥ 0.80)
source	string	Originating provenance marker

Usage

from datasets import load_dataset

train = load_dataset("agbalu/KabSentiment", split="train")
dev   = load_dataset("agbalu/KabSentiment", split="dev")
test  = load_dataset("agbalu/KabSentiment", split="test")

print(train[0])
# {'id': 'kab_sent_train_00000', 'text_kab': 'Ifukal-nwen ur sɛin ara lsas.',
#  'text_en': 'Your argument is unfounded.', 'label': 0, 'label_name': 'negative',
#  'confidence_score': 0.838, 'source': 'tatoeba_roberta_labeled'}

label encodes sentiment as 0 (negative), 1 (neutral), or 2 (positive), with label_name providing string identifiers. text_en provides the English parallel sentence from which the sentiment label was projected, included to enable auditability and cross-lingual analysis. confidence_score records the English teacher model's prediction probability.

Curation

Source Text

Sentences are human-written Kabyle text from the Tatoeba export (tatoeba_kab_eng_2026-08-05, 140,324 sentences). The export was deduplicated on the Kabyle side, filtered to sentence lengths between 4 and 25 words, and stripped of URLs, isolated digits, handle mentions, and currency symbols, yielding 109,723 candidate sentences.

Cross-Lingual Label Projection

Labels were projected from the aligned English source sentences using cardiffnlp/twitter-roberta-base-sentiment-latest (Loureiro et al., TimeLMs, 2022).

Predictions were retained only when teacher model confidence was $\ge 0.80$, filtering out 57% of candidates to eliminate ambiguous instances.

Candidate distribution following confidence thresholding:

Class	Retained Candidates	Subsampled for Release
Negative	8,189	5,000
Neutral	31,119	5,000
Positive	8,027	5,000

Each class was subsampled to exactly 5,000 sentences using seed 42.

Orthography

All text is normalized under the AƔBALU reference normaliser (1.3.0+rules1.0.0, 81 rules), preserving all standard phonemic diacritics (ɣ, ɛ, ḥ, ḍ, ṣ, ṭ, ẓ, ṛ, č, ǧ).

Baseline Benchmarks

Evaluated on test.parquet (1,500 sentences: 521 negative, 490 neutral, 489 positive):

System / Model	Setting	Accuracy	Macro-F1	Negative F1	Neutral F1	Positive F1
Masinissa-31M	Linear probe, frozen encoder	77.53%	0.7764	0.7389	0.8298	0.7604
Masinissa-31M	Full fine-tune	88.80%	0.8880	0.8831	0.9111	0.8697
Masinissa-Sentiment	Multi-seed average + calibrated	90.27%	0.9026	0.9010	0.9160	0.8910

The frozen encoder linear probe evaluates representations by fitting a single linear layer over mean-pooled token states (77.53% accuracy). Full fine-tuning unfreezes encoder parameters with a projection bottleneck, reaching 88.80% accuracy. Both models select checkpoints on the validation split via macro-F1 before evaluation on the unseen test split.

Benchmark Context

Prior sentiment resources for Kabyle were limited to binary classification without neutral labels, and frequently suffered from character loss where sub-dot diacritics were stripped during legacy font conversions. KabSentiment provides a balanced three-class dataset with canonical, verified Latin orthography.

Known Limitations

• Projected annotations: Labels originate from an English teacher classifier via parallel sentence pairs rather than native human sentiment tagging in Kabyle.
• Syntactic nuance: While literal sentiment transfers reliably, culturally specific idioms or ironic expressions may not be fully captured by cross-lingual projection.
• Shared sentence in train and dev: One sentence (Ulac ǧahennama yugaren ta.) appears as both kab_sent_train_06003 and kab_sent_dev_00065 due to distinct English source sentences translating to the identical Kabyle string. Train and test share zero strings.
• Register: The text represents written standard Kabyle in Latin script and does not cover dialectal spoken transcripts or heavy code-switching with French.

Citation

@misc{agbalu_kabsentiment,
  title  = {KabSentiment: a 3-class Kabyle sentiment benchmark},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabSentiment}
}

Please also cite the Tatoeba project for the source sentences, and Cardiff NLP for the teacher model (cardiffnlp/twitter-roberta-base-sentiment-latest).

Licence

CC-BY-4.0. The source Tatoeba sentences are CC-BY 2.0 FR; the compiled dataset is distributed under CC-BY-4.0 with attribution.