Datasets:
id stringlengths 20 20 | text_kab stringlengths 6 189 | text_en stringlengths 13 167 | label int64 0 2 | label_name stringclasses 3
values | confidence_score float64 0.8 0.99 | source stringclasses 1
value |
|---|---|---|---|---|---|---|
kab_sent_train_00000 | Ifukal-nwen ur sɛin ara lsas. | Your argument is unfounded. | 0 | negative | 0.838 | tatoeba_roberta_labeled |
kab_sent_train_00001 | Meslayeɣ d yelli. | I spoke to my daughter. | 1 | neutral | 0.897 | tatoeba_roberta_labeled |
kab_sent_train_00002 | D axeddam ifazen akk deg tkebbanit-nneɣ. | He is the best worker in our company. | 2 | positive | 0.984 | tatoeba_roberta_labeled |
kab_sent_train_00003 | Ttuɣ ur as-fkiɣ ara i weqjun-iw ad yečč. | I forgot to feed my dog. | 0 | negative | 0.8237 | tatoeba_roberta_labeled |
kab_sent_train_00004 | Rfan akk medden seg wayen i d-tennamt. | What you said made everyone angry. | 0 | negative | 0.9028 | tatoeba_roberta_labeled |
kab_sent_train_00005 | Usiɣ-d deg usakal. | I came by bus. | 1 | neutral | 0.8502 | tatoeba_roberta_labeled |
kab_sent_train_00006 | Ad yefṛeḥ aṭas imi i texdem ayenni. | He'll be very glad she did that. | 2 | positive | 0.9576 | tatoeba_roberta_labeled |
kab_sent_train_00007 | Nekṛeh-it i sin yid-nneɣ. | We both hate him. | 0 | negative | 0.886 | tatoeba_roberta_labeled |
kab_sent_train_00008 | Ẓriɣ Tom ixeddem aya. | I know Tom does that. | 1 | neutral | 0.8211 | tatoeba_roberta_labeled |
kab_sent_train_00009 | Fkant-ak-d kra ad teččeḍ-t? | Did they give you anything to eat? | 1 | neutral | 0.9359 | tatoeba_roberta_labeled |
kab_sent_train_00010 | Tella tin i yebɣan ad tceyyeɛ tabrat. | Somebody wants to send a message. | 1 | neutral | 0.8392 | tatoeba_roberta_labeled |
kab_sent_train_00011 | Tettbin-d d taneblalt tudert-im. | Your life seems perfect. | 2 | positive | 0.9216 | tatoeba_roberta_labeled |
kab_sent_train_00012 | Teqqim srid deffir-s. | She's sitting directly behind him. | 1 | neutral | 0.9216 | tatoeba_roberta_labeled |
kab_sent_train_00013 | Yesɛa mmi-s nniḍen. | He has another son. | 1 | neutral | 0.8859 | tatoeba_roberta_labeled |
kab_sent_train_00014 | Yella win i twalaḍ yeṭṭafar-ik-id? | Did you see anybody following you? | 1 | neutral | 0.9472 | tatoeba_roberta_labeled |
kab_sent_train_00015 | Mary tesɛa yemma-s i tt-iḥemmlen. | Mary has a mother who loves her. | 2 | positive | 0.9068 | tatoeba_roberta_labeled |
kab_sent_train_00016 | Targit-iw teqqel d tilawt. | My dream comes true. | 2 | positive | 0.9658 | tatoeba_roberta_labeled |
kab_sent_train_00017 | Beɛɛed ɣef yelli ! | Stay away from my daughter! | 0 | negative | 0.8322 | tatoeba_roberta_labeled |
kab_sent_train_00018 | Ḥemmleɣ mi ara tettgem aya. | I like it when you do that. | 2 | positive | 0.9271 | tatoeba_roberta_labeled |
kab_sent_train_00019 | Yekkat-d wedfel deg Boston? | Does it snow in Boston? | 1 | neutral | 0.902 | tatoeba_roberta_labeled |
kab_sent_train_00020 | Ur tezmireḍ ara i yiman-ik. | You are not able to care for yourself. | 0 | negative | 0.8395 | tatoeba_roberta_labeled |
kab_sent_train_00021 | Tom ha-t-an deg tkeṛṛust. | Tom is in the car. | 1 | neutral | 0.8927 | tatoeba_roberta_labeled |
kab_sent_train_00022 | Ur bɣiɣ ara ad zewǧeɣ akked Tom. | I don't want to marry Tom. | 0 | negative | 0.8286 | tatoeba_roberta_labeled |
kab_sent_train_00023 | Yeɛǧeb-iyi-d uqeṣṣer yid-k. | I've enjoyed talking to you. | 2 | positive | 0.971 | tatoeba_roberta_labeled |
kab_sent_train_00024 | La ken-tettraǧu tmes. | Hell is waiting for you. | 0 | negative | 0.8678 | tatoeba_roberta_labeled |
kab_sent_train_00025 | Ḥemleɣ-kem ugar i kem-iḥemmel Tom. | I love you more than Tom does. | 2 | positive | 0.9633 | tatoeba_roberta_labeled |
kab_sent_train_00026 | Feṛḥeɣ aṭas imi ad teqqimeḍ. | I'm so glad you're staying. | 2 | positive | 0.9841 | tatoeba_roberta_labeled |
kab_sent_train_00027 | Ssiwel-as i Tom, ini-as aql-aɣ-in. | Call Tom and tell him we're on our way. | 1 | neutral | 0.8567 | tatoeba_roberta_labeled |
kab_sent_train_00028 | Walaɣ belli la skiddibeɣ i yiman-iw. | I realised that I was lying to myself. | 0 | negative | 0.8199 | tatoeba_roberta_labeled |
kab_sent_train_00029 | Angul-inem, d aẓidan. | Your cake is delicious. | 2 | positive | 0.9746 | tatoeba_roberta_labeled |
kab_sent_train_00030 | Tella tsuqelt n tefṛansist? | Is there a French translation? | 1 | neutral | 0.9303 | tatoeba_roberta_labeled |
kab_sent_train_00031 | Mary ha-tt-an tezdeɣ d twacult-is. | Mary is living with her family. | 1 | neutral | 0.9266 | tatoeba_roberta_labeled |
kab_sent_train_00032 | Bɣiɣ ad ffɣeɣ yid-k. | I'd love to go out with you. | 2 | positive | 0.9647 | tatoeba_roberta_labeled |
kab_sent_train_00033 | Teswaɣ-iyi tudert-iw. | She ruined my life. | 0 | negative | 0.9387 | tatoeba_roberta_labeled |
kab_sent_train_00034 | Tatoeba d asmel igerrzen i ulmad n tutlayin nniḍen. | Tatoeba is a good website for language learning. | 2 | positive | 0.9294 | tatoeba_roberta_labeled |
kab_sent_train_00035 | Acḥal n yimdukkal uqriben i tesɛamt? | How many close friends do you have? | 1 | neutral | 0.9338 | tatoeba_roberta_labeled |
kab_sent_train_00036 | Ayweqt i ṛuḥen? | When did they go? | 1 | neutral | 0.9169 | tatoeba_roberta_labeled |
kab_sent_train_00037 | Bɣan-ken temmutem. | They want you dead. | 0 | negative | 0.855 | tatoeba_roberta_labeled |
kab_sent_train_00038 | S tidet iḥemmel-ik umcic-nneɣ. | Our cat really likes you. | 2 | positive | 0.9607 | tatoeba_roberta_labeled |
kab_sent_train_00039 | Werǧin ad cerheɣ. | I'll never be happy. | 0 | negative | 0.8719 | tatoeba_roberta_labeled |
kab_sent_train_00040 | Tkellxeḍ i iman-ik. | You're lying to yourself. | 0 | negative | 0.8088 | tatoeba_roberta_labeled |
kab_sent_train_00041 | Dayen ibanen, ad k-nɛawen. | Of course, we'll help you. | 2 | positive | 0.852 | tatoeba_roberta_labeled |
kab_sent_train_00042 | Ur tettkil ara, yak ? | She's unfriendly, isn't she? | 0 | negative | 0.9206 | tatoeba_roberta_labeled |
kab_sent_train_00043 | Acuɣer Ifransisen kerhen akk medden? | Why do the French hate everyone? | 0 | negative | 0.9066 | tatoeba_roberta_labeled |
kab_sent_train_00044 | Nḥemmel-it ur neẓri ayɣer. | We love him for reasons we don't fully understand. | 2 | positive | 0.8407 | tatoeba_roberta_labeled |
kab_sent_train_00045 | Ur zmireɣ ad xedmeɣ acemma deg aya. | I can do nothing about that. | 0 | negative | 0.8856 | tatoeba_roberta_labeled |
kab_sent_train_00046 | Yeɛreq cced i uyeddid! | It's really a mess forever! | 0 | negative | 0.9126 | tatoeba_roberta_labeled |
kab_sent_train_00047 | Tom akken d-yuɣal si Boston. | Tom just got back from Boston. | 1 | neutral | 0.9222 | tatoeba_roberta_labeled |
kab_sent_train_00048 | Sḥassfeɣ maḍi imi ur ssawḍeɣ ad ɛawneɣ. | I'm real sorry that I wasn't able to help. | 0 | negative | 0.8922 | tatoeba_roberta_labeled |
kab_sent_train_00049 | Wagi d ugur meqqren mliḥ. | This is a very big problem. | 0 | negative | 0.8723 | tatoeba_roberta_labeled |
kab_sent_train_00050 | Tesnemmer-iyi-d ɣef usefk-nni. | She thanked me for the present. | 2 | positive | 0.9554 | tatoeba_roberta_labeled |
kab_sent_train_00051 | Ur qebbleɣ ara axeddim-agi. | I wouldn't accept that job. | 0 | negative | 0.8951 | tatoeba_roberta_labeled |
kab_sent_train_00052 | Tenna-d dakken ur telli ara tezha. | She said that she wasn't enjoying herself. | 0 | negative | 0.8038 | tatoeba_roberta_labeled |
kab_sent_train_00053 | Mazal la d-tettbanemt terfamt. | You still look angry. | 0 | negative | 0.812 | tatoeba_roberta_labeled |
kab_sent_train_00054 | Ɣret idlisen, ad tlemdem aṭas n tɣawsiwin. | Read books, you'll learn many things. | 2 | positive | 0.8536 | tatoeba_roberta_labeled |
kab_sent_train_00055 | Yessaggad lḥal aṭas. | It is so worrisome. | 0 | negative | 0.8802 | tatoeba_roberta_labeled |
kab_sent_train_00056 | Ur yelhi ara lweqt i sɛeddaɣ ass-nni n lḥedd. | I didn't have a good time last Sunday. | 0 | negative | 0.9296 | tatoeba_roberta_labeled |
kab_sent_train_00057 | Ugin ad ḥebsen idammen. | It won't stop bleeding. | 0 | negative | 0.8368 | tatoeba_roberta_labeled |
kab_sent_train_00058 | Isekla n ufresɣu atnan deg ujuǧǧeg-nsen ummid. | The Peach Trees are in their full bloom. | 2 | positive | 0.9452 | tatoeba_roberta_labeled |
kab_sent_train_00059 | Ffeɣ tamurt, ad timɣureḍ. | Travel, and you will flourish. | 2 | positive | 0.9382 | tatoeba_roberta_labeled |
kab_sent_train_00060 | Tom yettmeslay d yiwen n umsaɣ. | Tom is talking to a customer. | 1 | neutral | 0.889 | tatoeba_roberta_labeled |
kab_sent_train_00061 | Ur iyi-iḥsib ara d mmi-s. | He does not treat me like his son. | 0 | negative | 0.8164 | tatoeba_roberta_labeled |
kab_sent_train_00062 | Anwa i wen-t-ixedmen? | Who did this for you? | 1 | neutral | 0.931 | tatoeba_roberta_labeled |
kab_sent_train_00063 | Tessardeḍ tamgerṭ-ik? | Did you wash your neck? | 1 | neutral | 0.8949 | tatoeba_roberta_labeled |
kab_sent_train_00064 | Feṛḥeɣ imi ken-ssneɣ. | I'm glad to know you. | 2 | positive | 0.9807 | tatoeba_roberta_labeled |
kab_sent_train_00065 | Ur telli ara tebɣa ad ternu ad tesɛeddi akud yid-s. | She didn't want to spend any more time with him. | 0 | negative | 0.8086 | tatoeba_roberta_labeled |
kab_sent_train_00066 | Ur zmiren ara ad ḍṣen fell-aneɣ. | They cannot laugh at us. | 0 | negative | 0.8733 | tatoeba_roberta_labeled |
kab_sent_train_00067 | James Allison akked Tasuku Honjo rebḥen arraz Nobel n tujjya. | James Allison and Tasuku Honjo won the Nobel prize in medicine. | 2 | positive | 0.822 | tatoeba_roberta_labeled |
kab_sent_train_00068 | Ma yella d ayen iɣeṣben, ssiwel-iyi-d s uṭṭun-a. | If it's urgent, call me at this number. | 1 | neutral | 0.8 | tatoeba_roberta_labeled |
kab_sent_train_00069 | Yettwajreḥ uqjun-nteɣ. | Our dog got hurt. | 0 | negative | 0.8665 | tatoeba_roberta_labeled |
kab_sent_train_00070 | Tirga-ik ad teffeɣ sya ɣer zdat kan. | Your dream will come true in the near future. | 2 | positive | 0.9658 | tatoeba_roberta_labeled |
kab_sent_train_00071 | Yettmeslay Tom Tafṛansist xiṛ n wakken i trujaḍ. | Tom speaks French better than you might expect. | 2 | positive | 0.851 | tatoeba_roberta_labeled |
kab_sent_train_00072 | Yuzzel wawal. | The word has spread. | 1 | neutral | 0.8259 | tatoeba_roberta_labeled |
kab_sent_train_00073 | D acu-tt tɣawsa taɣwalit akk ay teččam ? | What's the most unusual thing you've ever eaten? | 1 | neutral | 0.8824 | tatoeba_roberta_labeled |
kab_sent_train_00074 | Tebɣamt ad turaremt lkarṭa? | Would you like to play cards? | 1 | neutral | 0.906 | tatoeba_roberta_labeled |
kab_sent_train_00075 | Iɛǧeb-iyi wagi. Ad t-awiɣ. | I like this. I'll take it. | 2 | positive | 0.944 | tatoeba_roberta_labeled |
kab_sent_train_00076 | D tamdakkelt ɛzizen. | She is a dear friend. | 2 | positive | 0.9314 | tatoeba_roberta_labeled |
kab_sent_train_00077 | Tom d Mary llan ttemɣunzan. | Tom and Mary despised each other. | 0 | negative | 0.8508 | tatoeba_roberta_labeled |
kab_sent_train_00078 | Yesṛuḥ akk idrimen-nnes deg ukazinu. | He lost all of his money at the casino. | 0 | negative | 0.8277 | tatoeba_roberta_labeled |
kab_sent_train_00079 | Tuɛer tefṛansist i tɣuri. | French is difficult to read. | 0 | negative | 0.8593 | tatoeba_roberta_labeled |
kab_sent_train_00080 | Tom yeqqim ɣer yiri n tmes, la yesseḥmaw ifassen-is. | Tom stood by the fire, warming his hands. | 1 | neutral | 0.8698 | tatoeba_roberta_labeled |
kab_sent_train_00081 | Rfant mi asent-nniɣ akken. | They got angry when I told them that. | 0 | negative | 0.8347 | tatoeba_roberta_labeled |
kab_sent_train_00082 | Yesseḍṣ-it-id. | He made him laugh. | 2 | positive | 0.8664 | tatoeba_roberta_labeled |
kab_sent_train_00083 | Tom ad d-yas deg 20 Tubeṛ. | Tom will come on October 20th. | 1 | neutral | 0.9152 | tatoeba_roberta_labeled |
kab_sent_train_00084 | Almud ileddi tiwwura timaynutin. | Learning opens new doors. | 2 | positive | 0.8958 | tatoeba_roberta_labeled |
kab_sent_train_00085 | D ayyuren aya seg wasmi ay la tessefray mary ad tessikel. | Mary’s been planning her trip for months. | 1 | neutral | 0.8453 | tatoeba_roberta_labeled |
kab_sent_train_00086 | Tesɛiḍ arkasen d yiqaciren? | Do you have shoes and socks? | 1 | neutral | 0.9415 | tatoeba_roberta_labeled |
kab_sent_train_00087 | D tungift, maca zeddiget nneyya-s. | She's stupid, but honest. | 0 | negative | 0.8528 | tatoeba_roberta_labeled |
kab_sent_train_00088 | Tiɣawsiwin yelhan s drus i d-ttasent. | Good things come in small amounts. | 2 | positive | 0.968 | tatoeba_roberta_labeled |
kab_sent_train_00089 | Tessemɣaremt tamsalt. | You're exaggerating the problem. | 0 | negative | 0.8129 | tatoeba_roberta_labeled |
kab_sent_train_00090 | Ur d-yeqqim kra n usirem. | There is not much hope. | 0 | negative | 0.8033 | tatoeba_roberta_labeled |
kab_sent_train_00091 | Ur ttuɣaleɣ ara ad amneɣ isertanen. | I no longer have any confidence in politicians. | 0 | negative | 0.8813 | tatoeba_roberta_labeled |
kab_sent_train_00092 | Tejmeɛ akk tibṛatin-is. | She kept all his letters. | 1 | neutral | 0.8742 | tatoeba_roberta_labeled |
kab_sent_train_00093 | Tom ad yeqqim ɣer deffir. | Tom will sit in the back. | 1 | neutral | 0.8902 | tatoeba_roberta_labeled |
kab_sent_train_00094 | D imeslayen-ik i d-yeglan s reffu-s. | Your words provoked his anger. | 0 | negative | 0.8124 | tatoeba_roberta_labeled |
kab_sent_train_00095 | D anta i d tamazdayt-ik? | Who is your community? | 1 | neutral | 0.8714 | tatoeba_roberta_labeled |
kab_sent_train_00096 | Ur zmireɣ ad kent-wufqeɣ. | I cannot agree with you. | 0 | negative | 0.8761 | tatoeba_roberta_labeled |
kab_sent_train_00097 | Meqqer wayen iss tettekkiḍ. | You made a great contribution. | 2 | positive | 0.9664 | tatoeba_roberta_labeled |
kab_sent_train_00098 | D tidet telha, neɣ uhu ? | She's really cool, isn't she? | 2 | positive | 0.9827 | tatoeba_roberta_labeled |
kab_sent_train_00099 | Yezga yezzuzzur tikerkas. | He persistently spreads lies. | 0 | negative | 0.8397 | tatoeba_roberta_labeled |
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
trainanddev.Ulac ǧahennama yugaren ta.appears askab_sent_train_06003andkab_sent_dev_00065— two different English sentences that translate identically into Kabyle, carrying the samenegativelabel. 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 excludekab_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.
- Downloads last month
- 79