Granite-278m-kk — Kazakh retrieval embedding

A compact (278M) sentence-embedding model for Kazakh search and RAG, fine-tuned from ibm-granite/granite-embedding-278m-multilingual (R1). Kazakh is not in Granite's official language list — this model adapts it to Kazakh retrieval with targeted fine-tuning, while staying lightweight (~556 MB, fp16).

• 768-dim embeddings, cosine similarity, max sequence length 512, no prompt prefixes.
• Current revision is v2 (morphology-hardened). The previous model is kept as revision="v1".
• Full method, data, evaluation and scripts: github.com/Tim2190/kazakh-granite-retriever.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("Tim2190/granite-278m-kk")   # ~556 MB, fp16 (v2)
query = "Балқаш көлі қайда орналасқан?"
passages = ["Балқаш — Қазақстанның оңтүстік-шығысындағы тұйық көл ..."]
scores = model.similarity(model.encode([query]), model.encode(passages))
# previous version: SentenceTransformer("Tim2190/granite-278m-kk", revision="v1")

No special query/passage prefixes are needed. For stronger retrieval you can fuse the dense scores with a lexical channel (BM25 + a Kazakh stemmer) via Reciprocal Rank Fusion — see the eval scripts (eval.py --hybrid).

Training

• Data: 57,369 synthetic Kazakh (query → passage) pairs over Kazakh Wikipedia passages from KazQAD, each with a hard negative mined via BM25 with a Kazakh stemmer, plus 3,733 KazQAD labeled gold pairs (with their rel=0 negatives) — 61,102 training examples. Training data is de-duplicated against the evaluation benchmarks (anti-leak).
• Objective: CachedMultipleNegativesRankingLoss (sentence-transformers), 2 epochs, lr 1e-5, max_seq_len 512, Kaggle T4.
• Base: ibm-granite/granite-embedding-278m-multilingual (Apache-2.0).

Evaluation

Evaluated at seq 512 with the benchmark's own harness and paired-bootstrap significance (10k). Fine-tuning significantly improves the base Granite on Kazakh retrieval (in-domain nDCG@10 0.672 → 0.751 dense / 0.813 hybrid, p<0.001; out-of-domain speeches 0.430 → 0.529, all tiers p<0.05). The model is on par with the specialized kazakh-e5 (ALL: statistical tie, p=0.42), and the stemmer-mined hard negatives significantly improve morphological (inflected) queries over the previous v1 (p=0.002), closing the one slice where kazakh-e5 previously led. It does not beat kazakh-e5 overall. Full tables, baselines and significance tests are in the project repository.

License & attribution

Training passages come from KazQAD / Kazakh Wikipedia (CC BY-SA 4.0); the model is released under CC BY-SA 4.0 with attribution to KazQAD and Kazakh Wikipedia. The base Granite model is Apache-2.0.

Citation

Please cite the project repository and the underlying benchmarks/datasets (Kaz-RAG-search-benchmark, RAG-Two-Pass-Retrieval-QAZ, KazQAD).