KabEncBench: The Kabyle Representation Learning Benchmark

KabEncBench is the standard multi-task evaluation suite for contextual text representations and neural encoders in Kabyle (Taqbaylit, kab, Latin script).

Spanning 100,609 annotated items across six core linguistic tasks and three diagnostic probes, KabEncBench isolates what pretrained language models understand about Kabyle phonology, morphology, syntax, semantics, and discourse—without allowing deep task heads to mask representational failure.

Every task is evaluated under Track A: Frozen Scale-Normalized Probe, scoring models through a unified Bradley–Terry Elo tournament with 2,000 paired cluster-bootstraps anchored at FacebookAI/xlm-roberta-base = 1,000.

from datasets import load_dataset

# Load any of the 6 core tasks or 3 diagnostic suites
syntax = load_dataset("agbalu/KabEncBench", "syntax")
morph = load_dataset("agbalu/KabEncBench", "morph")
sentiment = load_dataset("agbalu/KabEncBench", "sent")
spell_diag = load_dataset("agbalu/KabEncBench", "diag_spell")

# Example inspection: Universal Dependencies syntax
print(syntax["test"][0]["tokens"])
print(syntax["test"][0]["upos"])

---

1. Official Bradley–Terry Leaderboard

Ratings derived from 2,000 paired cluster-bootstraps across all task test instances, anchored at FacebookAI/xlm-roberta-base = 1,000.0:

Rank	Model	Parameters	Elo Rating (95% CI)	Tier	Mean Probe Score
1	agbalu/Masinissa-73M	73M	1,953.7 [1,534.7, 3,379.1]	Tier 1	85.61%
2	agbalu/Masinissa-31M	31M	1,708.4 [1,371.1, 2,886.2]	Tier 2	83.78%
3	UBC-NLP/serengeti-E250	278M	1,200.3 [953.6, 1,480.1]	Tier 3	78.76%
4	Davlan/afro-xlmr-base-76L_script	278M	1,151.6 [919.5, 1,451.0]	Tier 3	77.74%
5	UBC-NLP/serengeti-E110	110M	1,148.2 [918.5, 1,421.2]	Tier 3	78.27%
6	jhu-clsp/mmBERT-base	278M	1,141.7 [906.4, 1,435.3]	Tier 3	77.36%
7	jhu-clsp/mmBERT-small	135M	1,039.1 [804.2, 1,282.5]	Tier 3	76.17%
8	google-bert/bert-base-multilingual-cased	178M	1,012.6 [780.9, 1,266.8]	Tier 4	77.02%
9	FacebookAI/xlm-roberta-base (Anchor)	278M	1,000.0 [1,000.0, 1,000.0]	Tier 4	76.87%

Task-by-Task Performance Matrix (%)

Model	UPOS (Syntax)	Aspect (Morph)	Micro-F1 (NER)	Sentiment	SIB-200 (Topic)	Punctuation	TaPaCo (Paraphrase)	Minimal Pairs (Spell)
agbalu/Masinissa-73M	80.59	57.91	88.24	93.83	64.87	55.42	64.38	60.90
agbalu/Masinissa-31M	78.88	54.19	87.74	92.06	59.28	52.21	58.78	18.70
serengeti-E250	76.98	49.13	83.98	85.33	43.72	45.54	42.55	6.30
serengeti-E110	76.24	48.83	84.91	85.65	39.64	43.11	43.86	4.70
mmBERT-base	75.52	48.44	84.01	85.64	37.96	39.98	48.63	53.30
mmBERT-small	75.72	45.84	81.79	83.17	34.94	39.91	48.98	61.00
xlm-roberta-base	75.44	44.52	79.71	84.75	38.21	39.12	42.13	57.40
afro-xlmr-base	75.32	47.93	83.04	85.75	38.41	38.56	48.54	62.50
bert-base-multilingual	73.21	46.79	80.62	84.35	40.84	39.01	51.77	27.00

---

2. Benchmark Architecture & Task Inventory

KabEncBench balances fundamental linguistic competencies across two tracks:

Core Probing Tasks (Track A)

1. syntax (Universal Dependencies UPOS & Syntax):
   Gold UD_Kabyle-ADPT treebank split (865 train, 220 dev, 845 test). Evaluates syntactic word-class identification over 16 universal POS tags. 29.5% of Kabyle words sit inside multiword tokens (MWTs); word representations are aligned via the initial subtoken.
2. morph (Factored Verbal Aspect & Mood):
   25,129 verbal predicates (21,160 train, 1,323 dev, 2,646 test) covering Aorist, Preterite, Intensive Preterite, and Participle forms. Morphology is factored to isolate aspectual semantics from lexical root identity.
3. ner (Named Entity Recognition):
   2,900 curated Kabyle sentences (2,000 train, 300 dev, 600 test) annotated for PER, LOC, ORG, and MISC entity spans under standard BIO tagging.
4. sent (Sentence Sentiment Classification):
   7,050 sentences (6,000 train, 450 dev, 600 test) balanced across 3 polarity classes (negative, neutral, positive). Curated with English-Kabyle confidence verification.
5. topic (SIB-200 7-Way Topic Categorization):
   1,004 sentences (701 train, 99 dev, 204 test) across 7 news topics (science/tech, sports, politics, etc.). Text repaired to canonical Kabyle Latin orthography, removing homoglyph corruptions present in public crawl datasets.
6. punct (Punctuation & Case Restoration):
   42,246 sequences (30,000 train, 5,052 dev, 4,694 test, 2,500 OOD). Evaluates contextual boundary detection across comma, period, question mark, and colon placement.

Diagnostic Probes (Track B)

7. diag_clitic (Clitic Host Attachment):
   3,241 sequences (1,471 train, 384 dev, 1,386 test). Diagnoses whether representations accurately identify pronominal and directional clitic boundaries (d/n particles, direct/indirect object clitics) following prepositions or verb stems.
8. diag_para (TaPaCo Zero-Shot Paraphrase Retrieval):
   15,919 sentences across 5,551 semantic clusters drawn from TaPaCo-kab. Backbones pool sentence embeddings via mean-pooling without fine-tuning; evaluated via Mean Average Precision at R (mAP@R) and Mean Reciprocal Rank (MRR@10).
9. diag_spell (Minimal Pairs Orthographic & Morphological Sensitivity):
   1,000 controlled minimal sentence pairs testing four structural linguistic phenomena:
   • SPIRANT_TH: Spirantized dental fricative contrast (ṯ vs. standard Latin t).
   • EMPHATIC_DROP: Loss of phonemic pharyngealization (ḍ, ḥ, ṛ, ṣ, ṭ, ẓ vs. plain consonants).
   • CONSTRUCT_ANNEX: Annexed construct state vocalic alternation (e.g. axxam $\to$ wexxam).
   • CLITIC_ORDER: Pre-verbal clitic climbing vs. ungrammatical post-verbal attachment in negative/relative clauses.

---

3. Evaluation Protocol

To prevent overfitting, hyperparameter hacking, or prompt engineering from distorting model capabilities, KabEncBench enforces strict evaluation invariants:

The Frozen Scale-Normalized Probe

• Frozen Backbone: The transformer weights $\theta$ receive zero gradient updates ($\nabla_\theta \mathcal{L} = 0$).
• ScalarMix Layer Weighting: Rather than arbitrarily evaluating the final layer, representations are computed via a learnable softmax over all $L+1$ layer representations (including embedding layer 0): $$\mathbf{h}i = \sum{l=0}^L \gamma_l \mathbf{h}_i^{(l)}, \quad \gamma_l = \frac{e^{\alpha_l}}{\sum_k e^{\alpha_k}}$$
• Linear Probe Head: A single linear layer projects pooled word/sentence representations to class logits: $$\hat{\mathbf{y}} = \mathbf{W} \mathbf{h} + \mathbf{b}$$
• Standardized Optimizer & Hyperparameter Grid: Probes are optimized with AdamW across a fixed grid of 5 learning rates ($[10^{-4}, 3\times 10^{-4}, 10^{-3}, 3\times 10^{-3}, 10^{-2}]$). For each learning rate, 5 random seeds are trained for 20 epochs. The learning rate achieving highest dev score is selected, and evaluated once on the test split.

Bradley–Terry Elo Tournament

For every bootstrap iteration $b \in {1, \dots, 2000}$, test instances are resampled with replacement within task clusters. Pairwise win/loss/tie outcomes are compiled across all model pairs, and maximum-likelihood latent skill parameters $\mu_m$ are estimated under the Bradley–Terry logistic model: $$P(\text{Model }A>\text{Model }B)=\frac{1}{1+{10}^{({\mu }_B-{\mu }_A)\mathrm{/}400}}P(\backslash text\{Model\; \}\; A\; >\; \backslash text\{Model\; \}\; B)\; =\; \backslash frac\{1\}\{1\; +\; 10^\{(\backslash mu\_B\; -\; \backslash mu\_A)\; /\; 400\}\}$$ Ratings are anchored at $\mu_{\text{XLM-RoBERTa-base}} \equiv 1,000.0$.

---

4. Decontamination & Scientific Rigor

• 4-Word Verbatim Decontamination: The test splits of all six core tasks and three diagnostics were scanned against the 34.9M token pretraining corpus of Masinissa. Any pretraining document sharing a 4-word verbatim sequence with an evaluation test sentence was scrubbed prior to pretraining.
• Positive Control Verification: A synthetic corpus containing known leak injections was evaluated through the decontamination harness to mathematically verify that the detector catches corruptions down to single homoglyph alterations.

---

5. Schema & Configuration Definitions

syntax

{
  "id": "kabsyntax_test_0042",
  "sent_id": "A104",
  "text": "Yerna-d yiwen n urgaz ɣer dinna.",
  "tokens": ["Yerna-d", "yiwen", "n", "urgaz", "ɣer", "dinna", "."],
  "upos": ["VERB", "NUM", "ADP", "NOUN", "ADP", "ADV", "PUNCT"],
  "head": [0, 1, 4, 2, 6, 1, 1],
  "deprel": ["root", "nsubj", "case", "nmod", "case", "advmod", "punct"]
}

diag_spell

{
  "id": "diag_spell_0112",
  "category": "CONSTRUCT_ANNEX",
  "good": "Iwala wexxam-nni seg lebɛid.",
  "bad": "Iwala axxam-nni seg lebɛid.",
  "lemma": "axxam"
}

---

6. Citation

@dataset{agbalu_kabencbench_2026,
  title  = {KabEncBench: A Multi-Task Representation Learning Benchmark for Kabyle},
  author = {AƔBALU},
  year   = {2026},
  url    = {https://huggingface.co/datasets/agbalu/KabEncBench},
  note   = {Covers 6 core tasks, 3 diagnostics, 100,609 annotated instances across Kabyle linguistics}
}