AI & ML interests

Natural Language Processing, Large Language Models, Algerian Darija, Arabic NLP, North African NLP, Multilingual NLP, Low-Resource Languages, Language Models, Dataset Curation, Evaluation

Recent Activity

ainouche-abderahmane  updated a Space 1 day ago
algerian-nlp/README
ainouche-abderahmane  updated a dataset 1 day ago
algerian-nlp/DziriAlign
ainouche-abderahmane  updated a dataset 1 day ago
algerian-nlp/DziriEval
View all activity

Organization Card

Algerian NLP Collective (algerian-nlp)

Open foundation models, speech systems, and open datasets for Algerian Darija (arq) across Arabic script (arq-Arab), Latin Arabizi (arq-Latn), French code-switching, and spoken audio.


Model Catalog

Model Task Architecture & Size Modality / Scripts License Headline / Notes
algerian-nlp/DZAIR Bidirectional Encoder 12L × 768 RTD + GDES (105.3M, 48k Unigram) Arabic, Arabizi, FR code-switch Apache-2.0 65.52% Narabizi Sentiment, 96.33% Ranim Sentiment; ships FP32, fp16/, onnx/ (FP32 & INT8)
algerian-nlp/BENKHLOUF Sentence Embeddings Matryoshka Bi-Encoder (278.1M, 768d → 64d) Arabic, Arabizi, FR/EN Apache-2.0 0.853 pos cosine (768d) / 0.897 (64d MRL), 0.708 Arabizi ↔ Arabic transfer
algerian-nlp/NUMALIA Maghrebi Dialect & Language ID Byte-Level ConvStem + 6L Transformer (5.19M, vocab 258) Raw UTF-8 bytes (Arabic, Arabizi, FR) CC-BY-SA-4.0 91.79% test acc / 91.59% macro F1 across 8 classes (DZ, MA, TN, LY, HS, MSA, FR, other_arabic) + 99% conformal sets
algerian-nlp/EL-ANKA Speech Recognition (ASR) Whisper-medium LoRA (69.2M adapter / 763.9M base) 16 kHz Algerian speech MIT 3-phase curriculum over 77,150 Algerian speech utterances (31,661 steps)
algerian-nlp/EL-ANKA-Lite Compact Speech Recognition (ASR) Whisper-small LoRA (26.0M adapter / 241.7M base) 16 kHz Algerian speech MIT Lightweight edge ASR adapter from the 3-phase Algerian speech curriculum (15,829 steps)
algerian-nlp/EL-HARRACHI Speech Synthesis (TTS) F5-TTS Flow-Matching DiT + Vocos (24 kHz) Arabic-script Darija + Latin FR Apache-2.0 Zero-shot voice cloning fine-tuned on ~399 h of Algerian speech (48,574 steps)
algerian-nlp/Hadra-TTS-kokoro Multi-Persona Speech Synthesis (TTS) Kokoro-82M / StyleTTS2 + iSTFTNet (24 kHz) Arabic-script Darija + FR loans Apache-2.0 Multi-speaker synthesis with three Algerian vocal personas (kahwa, rawi, loubna)

Datasets & Benchmarks

Dataset Scope & Size Type License
algerian-nlp/algerian-darja-corpus 11,151 long-form conversational transcripts (34.3M words) Pretraining Corpus CC-BY-4.0
algerian-nlp/algerian-arabic-english-translation-50k 50,000 aligned Algerian Darija–English pairs (Parquet, CSV, JSONL, JSON) Parallel Bitext Apache-2.0
algerian-nlp/Algerian-Youtube-Comments 55,365 raw Algeria-related YouTube comments (885 parquet shards) Social Corpus Unknown
algerian-nlp/TinyStories-Algerian-Darija 11,326 synthetic Darija stories paired with English originals (381 parquet shards) Synthetic Corpus Unknown
algerian-nlp/DziriEval 1,000 native Algerian Darija multiple-choice questions across 20 domains Evaluation Benchmark MIT
algerian-nlp/DziriAlign 1,000 Algerian Darija preference pairs (prompt, chosen, rejected) Alignment Dataset MIT

Quickstart

from transformers import AutoModel, AutoTokenizer
from sentence_transformers import SentenceTransformer

# 1. Contextual Encoder (FP32 default, or pass subfolder="fp16")
tok = AutoTokenizer.from_pretrained("algerian-nlp/DZAIR", trust_remote_code=True)
enc = AutoModel.from_pretrained("algerian-nlp/DZAIR", trust_remote_code=True).eval()

# 2. Cross-Script Sentence Embeddings (768d -> 64d Matryoshka)
embedder = SentenceTransformer("algerian-nlp/BENKHLOUF")
vecs = embedder.encode(["واش راك خويا", "wesh rak khoya"], normalize_embeddings=True)

Collective Standards

  • Cultural model naming: Model families carry Algerian cultural and historical names by task — DZAIR (encoders), BENKHLOUF (sentence embeddings), NUMALIA (Maghrebi dialect ID), EL-ANKA (ASR), and EL-HARRACHI (TTS) — using <FAMILY>[-<Gen>][-<Tier>] rather than upstream backbone tags.
  • Tri-script & code-switching fidelity: Arabic script, Latin Arabizi (with phoneme digits 3, 7, 9), and French code-switching are treated as first-class linguistic structure rather than noise to strip away.
  • Transparent provenance & licensing: Model cards separate weight/code licenses from underlying training data licenses, publish exact shipped file sizes and SHA-256 checksums, and state plainly what has been measured vs. what remains unmeasured.
  • Reproducible evaluation: Downstream comparisons score our models and baselines under matched protocols and splits, reporting deficits alongside wins.

License

Model weights and code are released under open licenses (typically Apache-2.0, MIT, or CC-BY-SA-4.0 as declared on each repository). Those grants apply to the published weights and code and do not relicense underlying third-party training text or audio.