DZAIR-ONNX

ONNX export (opset 18) of DZAIR, the 105.3M-parameter replaced-token-detection encoder for Algerian Darija — for CPU inference without PyTorch, via optimum[onnxruntime].

Numerically transparent: cosine similarity 1.0 against the torch outputs with max absolute difference 5.7e-06, accepted by the release fidelity gate. Takes input_ids plus attention_mask, returns last_hidden_state. Wrap chunks as [CLS] ... [SEP] and lowercase Latin spans first, exactly as with the base model. All results from the main card carry over: Arabizi sentiment 65.52, forum sentiment 96.33, both 10-seed means.

Usage

from optimum.onnxruntime import ORTModelForFeatureExtraction
from transformers import AutoTokenizer

REPO = "ainouche-abderahmane/DZAIR-ONNX"
tokenizer = AutoTokenizer.from_pretrained(REPO, trust_remote_code=True)
encoder = ORTModelForFeatureExtraction.from_pretrained(REPO)

For the full results table, training data, and licence composition, see the main card.

Files

file	bytes	contents
model.onnx	423,700,767	fp32 ONNX graph, opset 18
config.json	1,037	architecture and training summary
tokenizer.model	967,834	original SentencePiece model
tokenizer_config.json	370	fast-tokenizer wiring
tokenizer_rules.yaml	2,058	normalisation rules, versioned

Licence

Apache-2.0 for the weights and code, same grant and same training-text caveat as the base model — see the main card's licence composition before redistributing derivatives.