DZAIR-FP16

Half-precision build of DZAIR, the 105.3M-parameter replaced-token-detection encoder for Algerian Darija. Same weights, same tokenizer, half the bytes: 210,619,440 on disk against 421,228,128 for fp32.

Conversion changes nothing measurable: cosine similarity 1.00001 against the fp32 outputs with max absolute difference 0.0039, accepted by the release fidelity gate. All figures from the main model card carry over unchanged — Arabizi sentiment 65.52, forum sentiment 96.33, both 10-seed means.

Usage

Drop-in replacement for the fp32 repo at half the memory:

import torch
from transformers import AutoModel, AutoTokenizer

REPO = "ainouche-abderahmane/DZAIR-FP16"
tokenizer = AutoTokenizer.from_pretrained(REPO, trust_remote_code=True)
encoder = AutoModel.from_pretrained(
    REPO, trust_remote_code=True, torch_dtype=torch.float16,
).eval()

Lowercase Latin input first, as with the base model. For the full results table, training data, and licence composition, see the main card.

Files

file	bytes	contents
model.safetensors	210,619,440	fp16 discriminator weights
config.json	1,037	architecture and training summary
modeling_dzair.py	58,179	the architecture in code
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.