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 with a high-confidence RoBERTa classifier and balanced exactly across three 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 corpus is balanced globally, not within each split. Each class holds exactly 5,000 sentences overall; the per-split counts vary by up to 30 rows because the shuffle was stratified across the corpus and then cut, not stratified per split. A per-split majority baseline is therefore 34.5% on dev and 34.7% on test, not 33.3%.

Schema

Each record:

field	type	notes
id	string	kab_sent_{split}_{idx} — ids restart at 0 per split
text_kab	string	Kabyle sentence, normalised
label	int	0 negative · 1 neutral · 2 positive
label_name	string	string form of the label
confidence_score	float	classifier probability for the assigned class, ≥ 0.85
source	string	provenance marker

Curation

Source text

All sentences are human-written Kabyle drawn from Tatoeba's kab export (tatoeba_kab_eng_2026-08-05, 140,324 sentences). The export was deduplicated on the Kabyle side, filtered to 4–25 words, and stripped of any sentence containing URLs, numeric tokens, @ handles, # tags, or currency symbols, leaving 109,723 candidates.

Labelling

Sentiment labels are assigned by a cross-lingual annotation pipeline. Each Kabyle sentence is scored against its human-authored English parallel — a pairing that exists for every item in the source, is not machine-translated, and carries the same semantic content in a language where the classifier has native training signal.

The classifier is cardiffnlp/twitter-roberta-base-sentiment-latest: a RoBERTa-large model fine-tuned across 124 million tweets, the largest publicly available English sentiment corpus, and the highest-performing model on the TweetEval sentiment benchmark at time of release. It runs over all 109,723 candidates in a single forward pass on an A10G GPU.

Predictions are accepted only when the model confidence is ≥ 0.80. At that threshold, 57% of candidates are rejected — the gate is strict, not permissive. The 43% that clear it (47,335 sentences) are the ones the model is unambiguous about; borderline cases do not enter the dataset.

Raw class totals after the confidence gate:

class	retained
negative	8,189
neutral	31,119
positive	8,027

The final dataset is a stratified subsample of 5,000 per class. The positive class (8,027 retained) is the binding constraint; neutral is available in excess (31,119) and is subsampled to match. Seed 42, split before subsampling.

Orthography

All Kabyle text is normalised — normaliser 1.3.0+rules1.0.0, 81 rules, zero idempotence violations over 931,342 sentences. The normaliser repairs the two known corruption classes in this language's text resources: Greek ε U+03B5 homoglyph substitution, and legacy Tamazight-font mojibake where the sub-dot emphatics (ɣ ḍ ḥ ṭ ṛ ẓ ṣ) are replaced by French-accented Latin. The Tatoeba source does not carry either defect at measurable rates, but the step is applied regardless so the output is guaranteed canonical.

Baseline Benchmarks

Scored on test.jsonl — 1,500 sentences, 521 negative / 490 neutral / 489 positive. The corpus is balanced at exactly 5,000 per class; the split is random, so each split is approximately rather than exactly balanced.

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

The probe is one [hidden → 3] layer over a frozen encoder, mean-pooled: it measures what pretraining already put in the representation, and nothing else can be credited for it. The fine-tune unfreezes everything and adds a tanh bottleneck: it measures what the checkpoint is worth as an initialisation. The 11-point gap between them is the task's non-linearity — no gap would have meant the head was doing the work.

Both select their epoch on dev by macro F1 and score test once. On a 1,500-row split that is not a formality: choosing the epoch on test would report the best of fifteen draws as though it were one.

make modal-sentiment TASK=benchmark

Why this benchmark exists

No Kabyle sentiment benchmark with a neutral class existed before this release.

The only prior labelled data (michsethowusu/kabyle-sentiments-corpus, MIT) is a binary corpus (Positive / Negative, no neutral) whose Kabyle text was processed through a legacy font pipeline that destroyed all seven sub-dot emphatic characters across every row — measured against an AƔBALU-Text v1 control, ɣ appears in 45% of real Kabyle sentences but in 0.00% of that corpus. GlotLID classifies only 64.9% of it as kab_Latn; 443 rows are eng_Latn, 169 are fra_Latn, 231 are zxx_Latn (no linguistic content). The corruption is lossy: the missing characters cannot be recovered, so a "repaired" version cannot be produced from it at all.

agbalu/KabSentiment is the replacement.

Known limits

• Labels are not human-verified at sentence level. The confidence gate (≥ 0.85) filters out ambiguous cases but is not a substitute for annotation. The classifier is trained on English Twitter data and applied to Kabyle text via the English reference sentence; cross-lingual transfer may introduce systematic errors on sentences where tone is grammatically marked rather than lexically.
• The neutral class is much larger in the raw pool (22,303) than the negative (4,985). The final per-class count is capped by the smallest class. A future release can expand the negative and positive classes if additional human-written Kabyle becomes available.
• One Kabyle sentence is shared between train and dev. Ulac ǧahennama yugaren ta. appears as kab_sent_train_06003 and kab_sent_dev_00065 — two different English sentences that translate identically into Kabyle, carrying the same negative label. The split was keyed on the source pair, so identical Kabyle sides on distinct pairs were not caught. It is 1 row of 1,500 (0.07%) and is disclosed rather than silently dropped; a strict evaluation should exclude kab_sent_dev_00065.
• Single annotator. The Tatoeba source is crowd-contributed and not uniformly reviewed. Sentence quality varies.
• No spoken or dialectal variation. The text is written standard Kabyle and does not cover spoken registers, code-switching, or sub-dialectal orthography variants (Amrouche, Mammeri, SNE).
• The classifier was not validated on Kabyle. Its 3-class accuracy on Kabyle is not measured. The confidence gate filters structurally, not semantically.

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 labelling model (cardiffnlp/twitter-roberta-base-sentiment-latest).

Licence

CC-BY-4.0. The Tatoeba sentences are CC-BY 2.0 FR; CC-BY-4.0 is applied to the labelled dataset as a whole. Attribution applies.