CAMeLBERT-da-sentiment — Apple Core AI, Arabic dialect sentiment

A Core AI (.aimodel) build of CAMeL-Lab/bert-base-arabic-camelbert-da-sentiment for on-device dialectal Arabic sentiment on Apple Silicon.

Upstream ships pytorch_model.bin, flax_model.msgpack and tf_model.h5 — and no MLX, no ONNX, no Core AI. 81,297 downloads, and this is the Core AI build.

This is the dialect sibling of CAMeLBERT-mix-sentiment-CoreAI. Same three labels, trained on dialectal Arabic — and on the sentence below it disagrees with the MSA model, correctly.

Outputs

One graph, two outputs. Keeping both costs nothing at this size and means neither the trained classifier nor the reusable vectors are thrown away.

Read this before reading the outputs

Select outputs by shape, not by name. The runtime names tensors after the op that produced them. On this graph that is div_1 for the vectors and addmm_73 for the logits — the names in the wrapper's dict do not survive into the .aimodel. A one-line size == 3 / size == 768 pick is the robust way, and it is what the measurement script does.

Verified equivalence

Compared against the original PyTorch model on six fixed inputs at the exported shape:

Worst across all six: 4.53e-06, and 6/6 labels agree with PyTorch. That residual is relative ~1e-6 on logits of magnitude ~3 — float32 rounding accumulated through the classifier layer, not a difference in the model.

One dialect example, and what it does not show

الطلب اتأخر أسبوعين وما حد رد عليا. — an Egyptian-Arabic complaint — is classified negative here. The MSA model calls the same sentence neutral, which is the wrong answer for a complaint.

This is one sentence. It is the reason both bundles exist, but it is not a benchmark: no dialect sentiment corpus was run, and no accuracy figure for either model is claimed here.

Verified behaviour on six inputs, all matching PyTorch:

Use

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

LABELS = ["positive", "negative", "neutral"]
SEQUENCE_LENGTH = 512          # fixed shape: pad to it

async def analyse(text: str):
    tok = AutoTokenizer.from_pretrained(".")
    model = await AIModel.load("camelbert-arabic-sentiment.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"))})
    flat = {k: np.asarray(v.numpy()).reshape(-1) for k, v in out.items()}
    logits = next(v for v in flat.values() if v.size == 3)
    vectors = next(v for v in flat.values() if v.size == 768)
    return LABELS[int(np.argmax(logits))], logits, vectors

print(asyncio.run(analyse("الخدمة كانت ممتازة وأنصح به بشدة."))[0])   # positive

What is in here

Limits

• Fixed at 512 tokens; pad rather than passing a shorter sequence.
• Six inputs measured. Dialect variation, long documents and the 512-token boundary were not.
• No speed or memory benchmark. Equivalence only.
• The classification is 3-way and single-label; no confidence calibration was measured.

Licence and attribution

Apache-2.0, as declared by upstream. A derivative work of CAMeL-Lab/bert-base-arabic-camelbert-da-sentiment; credit for the model is theirs.