GATE-AraBert-v1 — Apple Core AI, Arabic sentence embeddings

A Core AI (.aimodel) build of Omartificial-Intelligence-Space/GATE-AraBert-v1 for on-device Arabic sentence embeddings on Apple Silicon.

Upstream ships model.safetensors and a sentence-transformers config — and no MLX, no ONNX and no Core AI. 14,743 downloads, and this is the Core AI build.

Read this first: the vectors are deliberately NOT normalised

Upstream's modules.json declares Transformer -> Pooling and nothing else, and 1_Pooling/config.json sets pooling_mode_mean_tokens: true. There is no 2_Normalize module — so the library returns mean-pooled vectors with no L2 normalisation, and so does this bundle. The norms below are ~21–23, not 1.0.

That is a deliberate difference from Vela-1.0-Encoder-307M-Embedding-CoreAI, whose upstream pipeline does include a Normalize module and whose bundle therefore normalises. Copy the wrong behaviour across bundles and the cosine scores stop matching the library.

If you want cosine similarity, normalise on your side:

v = v / np.linalg.norm(v)

Illustrative comparison with a masked-language BERT

GATE-AraBert was trained for similarity, and it shows. Compare with the CAMeLBERT-mix bundle, where mean pooling on a model with no such objective left unrelated Arabic sentences at 0.65:

Lower absolute scores but a wider margin on these pairs: 0.31 to 0.55 of separation against 0.16 to 0.21. The gap, not the height, is what matters for retrieval — but this is three sentence pairs, not an information-retrieval benchmark. No NDCG, MRR or MAP was computed.

Cross-lingual: "How do I reset my password?" sits at 0.4274 from its Arabic counterpart.

Verified equivalence

Measured against the library's own encoder — sentence_transformers.SentenceTransformer.encode with normalize_embeddings=False, not a reimplementation of its pooling:

Worst: 3.99e-06 on vectors of norm ~22 — relative ~1.8e-7, float32 rounding.

Use

import asyncio, numpy as np
from coreai.runtime import AIModel, NDArray, SpecializationOptions
from transformers import AutoTokenizer

SEQUENCE_LENGTH = 512          # fixed shape: pad to it

async def embed(text: str):
    tok = AutoTokenizer.from_pretrained(".")
    model = await AIModel.load("gate-arabert.aimodel", SpecializationOptions.default())
    fn = model.load_function("main")
    enc = tok([text], padding="max_length", truncation=True, max_length=SEQUENCE_LENGTH,
              return_tensors="np")
    out = await fn({"input_ids": NDArray(enc["input_ids"].astype("int32")),
                    "attention_mask": NDArray(enc["attention_mask"].astype("int32"))})
    v = np.asarray(list(out.values())[0].numpy()).reshape(-1).astype("float64")
    return v / np.linalg.norm(v)          # normalise yourself; upstream does not

print(asyncio.run(embed("كيف أستعيد كلمة المرور؟")).shape)   # (768,)

What is in here

Limits

• Fixed at 512 tokens; pad rather than passing a shorter sequence.
• Six inputs measured. Arabic dialects, long documents, and the 512-token boundary were not.
• No speed or memory benchmark. Equivalence only.
• Un-normalised output is intentional — see above. Do not assume cosine works without normalising.

Licence and attribution

Apache-2.0, as declared by upstream. A derivative work of Omartificial-Intelligence-Space/GATE-AraBert-v1; credit for the model is theirs.

One caveat, because the chain matters: GATE-AraBert-v1's card names Omartificial-Intelligence-Space/Arabic-Triplet-Matryoshka-V2 as its base, and the licence situation further up that chain is not declared. This bundle reproduces the licence that the model it redistributes declares, and makes no claim about anything deeper.