A 278M-parameter sentence transformer for Kabyle (Taqbaylit, kab, Latin script), fine-tuned from multilingual-e5-base on 511,549 decontaminated bitext pairs with Matryoshka Representation Learning.

It retrieves the correct Kabyle passage at Recall@1 = 97.0% against 499 in-batch distractors — +33.2 absolute points over its own backbone, which collapses on Kabyle — and its 64-d Matryoshka slice loses nothing measurable against the full 768-d vector.

It is, as far as we can establish, the first sentence embedding model for Kabyle evaluated against hard bitext distractors on a decontaminated, cluster-sealed split.

Named after Si Mohand ou Mhand (1848–1905), the supreme master of the Kabyle asefru — whose verses travel by ear and survive only through the fidelity of recall.

Bitext retrieval: 500 held-out pairs, one correct passage per query, 499 distractors. Every query is encoded once; its correct passage must rank first (Recall@1) or in the top 5 (Recall@5) among all 500 passages. MRR is the mean reciprocal rank over the same pool. All three models are run in the same GPU container under the same normalization policy.

Three things worth reading carefully.

The backbone's failure is architectural, not random. multilingual-e5-base achieves mean_pos = 0.8547 but mean_neg = 0.7680 — a margin of only +0.087. Every Kabyle sentence looks roughly the same to it because its XLM-RoBERTa backbone maps Ɛ, Ɣ, Ǧ, Ẓ, ẓ to <unk>, collapsing distinct consonants into the same token. The isotropy measurement confirms it: mean_cosine = 0.8040 on unaligned random pairs, against a collapse threshold of 0.50. The space is cone-shaped. SiMohand's isotropy is 0.0042 on the same measurement.

LaBSE scores 22.2% for a structural reason. Kabyle is not in LaBSE's 109 training languages. Its WordPiece tokenizer has no coverage for Kabyle morphology, so it encodes every word as a sequence of byte-level fallbacks, and the semantic signal is lost before the encoder sees it.

The distractor pool is harder than it looks. All 500 passages are Kabyle (or English/French translations of Kabyle sentences). The model is not separating Kabyle from English — it is finding the single correct Kabyle paraphrase or translation in a pool of 499 other valid Kabyle sentences from the same domain. The backbone's 63.8% is its ceiling on this corpus, not on easy cross-lingual retrieval.

SiMohand is trained with a 5-tier nested loss over [768, 512, 256, 128, 64] dimensions. Truncating to 64 dimensions — a 12× compression — loses no measurable Recall@1:

The 256-d and 128-d slices score higher than the full vector on Recall@1. This is not an anomaly: shorter slices push the loss harder on fewer coordinates, concentrating the discriminative signal. The margin increases monotonically as dimension falls. For applications where vector storage or latency is the binding constraint, 64 dimensions is the recommended operating point.

Semantic search, retrieval-augmented generation, sentence clustering, and cross-lingual alignment for Kabyle–English and Kabyle–French.

Not suitable for: generation, translation, or any language other than Kabyle (and English or French as the passage side of a cross-lingual retrieval pair). The model has not been evaluated for bias or toxicity. Its training corpus contains no paraphrase annotation for Kabyle beyond the TaPaCo clusters used in split construction — the translation_* pairs are bitext, not paraphrases, and retrieval of near-paraphrases within Kabyle has not been benchmarked. No safety evaluation of any kind has been performed.

sentence-transformers and torch:

Read the second column: the English translation of query 1 scores 0.853, above every Kabyle sentence that is not its paraphrase. The space is shared, so a Kabyle query retrieves across languages without a translation step.

At 64 dimensions — a 12× compression, and the same retrieval accuracy:

Truncate, then normalise — in that order. A slice of an already-normalised vector is not a unit vector, and cosine similarity computed from one is not cosine similarity. normalize_embeddings=True is what does it for the full 768; for a slice it is the normalize call above.

Every number in both blocks was executed against the published weights before being written down.

Fine-tuned from intfloat/multilingual-e5-base: 12 layers, 768 hidden, 12 heads, XLM-RoBERTa backbone with a mean-pooling head and an L2-normalisation step.

Before training, the vocabulary is expanded from 250,002 to 250,007 tokens. Five Kabyle consonants are missing from XLM-RoBERTa's SentencePiece vocabulary and map to <unk>:

Donor initialisation is phonologically grounded: ẓ takes voicing from z and emphasis from ṣ. ǧ takes the affricate from d and j, the decomposition the project normaliser already records for it. The repair is verified rather than assumed: assert_covered re-runs the coverage check after each addition and raises rather than warning if any character still routes through <unk>.

ẓ occurs at 3.94% of tokens in a 5,000-row AƔBALU-Text control. An encoder that silently drops it trains to a healthy loss while mapping aẓar and aar to the same representation. boffire/kabyle-sentence-transformer-mpnet, the only other published Kabyle sentence model, inherits this defect: it reports mean cosine similarity, which cannot detect it.

The loss is MatryoshkaLoss over MultipleNegativesRankingLoss. MNRL treats every off-diagonal passage in a batch of size 64 as an in-batch negative — 63 negatives per query per step. The Matryoshka wrapper computes the same InfoNCE objective five times at nested prefix slices and sums. The scale factor of 20.0 sharpens the softmax, which is standard for MNRL on this class of task.

Validation improved monotonically across all three epoch checkpoints:

The final post-training check over 1,000 training-set queries gives mean_cosine = 0.1208, std = 0.1612 — consistent with a healthy, uncollapsed embedding space.

All pairs are human-authored bitext from public parallel corpora — 18 sources in total. NLLB-mined bitext is excluded (training a fine-tune on the base backbone's own mined output distils its errors, not its knowledge). The TaPaCo paraphrase table for Kabyle is excluded because its licence composition is unresolved.

511,549 pairs: kab–eng 291,318 / kab–fra 220,231.

In-batch negatives assume semantically distinct rows. Parallel corpora group naturally into clusters (a source sentence paired with its English translation and its French translation share the same semantic content). Standard random batching places cluster siblings in the same batch with non-negligible probability, producing false-negative gradient penalties that fragment semantic clusters rather than align them.

ClusterAwareBatchSampler guarantees at most one item per cluster_id per mini-batch, eliminating false negatives without reducing batch diversity. With 511,549 pairs and 511,549 unique clusters (every bitext pair in its own cluster), the sampler provides the strictest possible guarantee. Epoch coverage is complete; no pair is dropped.

The dev split is constructed by split_clusters at seed 42, assigning 500 complete clusters to evaluation before any training data is seen. No cluster spans train and dev: because every pair has a unique cluster_id, 500 pairs are in dev and 511,549 are in train, with zero overlap by construction.

Benchmark decontamination — verifying that no dev sentence appeared in training — is implicit in the cluster seal. A pair that is in dev is in dev because its cluster_id was drawn first. The 500 dev clusters come from 12 distinct sources, covering both translation_eng (286 pairs) and translation_fra (214 pairs), matching the training distribution.

No FLORES+ or SIB-200 sentence appears in the training corpus by the same NFKD fingerprint check applied to Masinissa-31M.

One task, one domain, one harness. The 97.0% Recall@1 is measured on held-out bitext pairs from the same distribution as the training corpus. Kabyle–Kabyle paraphrase retrieval — finding a paraphrase written entirely in Kabyle — has not been benchmarked, because no annotated paraphrase test set for Kabyle exists. The TaPaCo clusters are the only public source and their licence is unresolved, so they are excluded from both training and evaluation. Treat the 97.0% as a bitext-retrieval number, not a general semantic similarity number.

The training corpus is bitext only. 511,549 pairs are kab–eng or kab–fra translations. No kab–kab paraphrase pairs enter training (TaPaCo is excluded). The model can retrieve an English or French passage given a Kabyle query, and vice versa. Its behaviour on monolingual Kabyle queries paired with Kabyle passages comes from implicit transfer and has not been measured against a gold standard.

The held-out pairs come from 12 of the 18 training sources. Dev pairs were drawn from the same source pool as training pairs, so any bias in the training sources — domain, register, speaker — also affects the evaluation. No out-of-domain split exists.

Sibling-language contamination is bounded, not cleared. The source corpora carry kab_Latn labels, but LID systems cannot reliably distinguish Kabyle from Tarifit, Central Atlas Tamazight or Shawiya (see Masinissa-31M for the measurement). The same contamination bound from the encoder pretraining applies here.

No human adequacy evaluation. The training pairs are assumed correct — a source sentence and its translation are treated as semantically equivalent. Adequacy errors, lexical errors or orthographic mismatches in any source pair propagate undetected. The 0.7913 mean positive cosine is an upper bound on training-pair quality, not a precision figure.

No safety evaluation of any kind has been performed.

This is the retrieval backbone for the AƔBALU NLP stack:

The training checkpoint, with AdamW moments and RNG state, is not published. Ask if you need it to resume training from epoch 3.

The H2H evaluation runs in the same container as training, loading all three models on the same GPU under the same normalization, and writes h2h_report.json to the model directory on the volume. The numbers in this card come from that file.

Si Mohand ou Mhand (c. 1848–1905) is the greatest poet of the Kabyle oral tradition — ameddaḥ by vocation, carrying the whole grammar of feeling in verse. He composed in asefru, the short lyric form, and never wrote down a word. Everything that survives was recovered from memory by those who had heard him, recalled line by line.

A model that retrieves meaning by fidelity of embedding — that must hold the semantic content of a sentence in a fixed-dimensional vector and recover it against 499 alternatives — does, in its technical way, what Si Mohand's listeners did: carry a representation that survives at distance.

His verse circulates without a canonical text. Its persistence is the persistence of representation. The naming is homage; it implies no endorsement by anyone.

Apache-2.0 for the weights. That grant does not relicense the text they were trained on. The training corpus is drawn from public parallel corpora with mixed licence composition, including permissive (MIT, CC-BY), share-alike (CC-BY-SA), and unclear or unresolved licences. Weights are Apache-2.0; text keeps its upstream licence.