CAMeLBERT-mix-ner-CoreAI — Apple Core AI

A Core AI (.aimodel) build of CAMeL-Lab/bert-base-arabic-camelbert-mix-ner for on-device Arabic inference on Apple Silicon.

Upstream ships PyTorch (and often ONNX/TF) weights and no Core AI build. This is one.

Outputs

Labels, in config.json id2label order:

0 B-LOC
1 B-MISC
2 B-ORG
3 B-PERS
4 I-LOC
5 I-MISC
6 I-ORG
7 I-PERS
8 O

Verified equivalence

Every number below was produced by running the exported graph and the original PyTorch model on the same inputs and diffing the raw outputs. The reference outputs are in ref-output.json, so the check can be re-run rather than believed.

Worst absolute difference 1.192e-05, against outputs that run up to 10.3 — relative 1.2e-06, i.e. float32 rounding. Minimum cosine 1.0000000000 between the embeddings.

This is arithmetic equivalence only. There is no quality benchmark here — no F1, no accuracy score, no comparison with upstream's reported numbers.

Use

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

LABELS = ['B-LOC', 'B-MISC', 'B-ORG', 'B-PERS', 'I-LOC', 'I-MISC', 'I-ORG', 'I-PERS', 'O']
SEQUENCE_LENGTH = 512      # fixed shape: pad to it

async def run(text: str):
    tok = AutoTokenizer.from_pretrained(".")
    model = await AIModel.load("camelbert-arabic-ner-mix.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"))})
    # the runtime names outputs after the op that produced them, so select by shape, not by name
    return [np.asarray(v.numpy()) for v in out.values()]

print([a.shape for a in asyncio.run(run("نص عربي"))])

What is in here

Limits

• Fixed at 512 tokens. Longer inputs need chunking.
• 8 sentences measured, listed above. Dialect text, long documents and domain-specific vocabulary were not.
• No quality benchmark and no speed or memory measurement.
• The graph returns raw outputs; pooling, softmax, argmax and BIO decoding are the caller's steps, exactly as in the library.

Licence and attribution

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