Emhotob-5M-Darija-MSA-v1

Emhotob-5M-Darija-MSA-v1 — Bidirectional Darija ↔ MSA (~5M params)

A 5.08M-parameter model that translates both ways between Moroccan Darija (الدارجة المغربية) and Modern Standard Arabic (الفصحى). A single set of weights serves both directions; a direction-specific Arabic system prompt selects which way to translate.

Finetuned from oddadmix/Emhotob-5M, a tiny Llama-architecture base (hidden size 128, 5 layers, 4 heads, tied embeddings).

Scaling study. This runs the exact recipe of oddadmix/50M-Darija-MSA-v1 on a base ~10× smaller. At 5M the model switches script/register correctly and is the strongest of the 5M translation set (MSA→Darija BLEU ~13), though it still drifts on longer inputs. A scaling demonstration, not a production translator; use the 50M sibling for fluent output.

Evaluation

Evaluated on a deterministic held-out set of ~3,000 pairs (seed=42; 2,961 scored after dropping rows with an empty MSA reference), decoded greedily (do_sample=False, no repetition penalty), scored with sacreBLEU:

MSA→Darija scores highest of any direction in the 5M suite. The saved weights are the best checkpoint by validation loss (eval_loss = 3.512, epoch 3 of 3). For reference, the 50M sibling scores BLEU ~32 (dar→msa) / ~40 (msa→dar) with the same data and eval.

Example translations

Real greedy-decoded outputs from the held-out set (rough — this is a 5M model):

Darija → MSA

MSA → Darija

The model produces authentic Darija markers (غادي, كانضن, هادشي) in the MSA→Darija direction and MSA morphology the other way, but frequently loses source content. 20 samples per direction with references are in eval_bidirectional.json.

Usage

ChatML format. Pick the system prompt for the direction you want:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "oddadmix/Emhotob-5M-Darija-MSA-v1"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16).to("cuda").eval()

SYS_TO_MSA = "أنت مترجم محترف. ترجم النص من الدارجة المغربية إلى اللغة العربية الفصحى."
SYS_TO_DAR = "أنت مترجم محترف. ترجم النص من اللغة العربية الفصحى إلى الدارجة المغربية."

def translate(text: str, system: str) -> str:
    prompt = (
        f"<|im_start|>system\n{system}<|im_end|>\n"
        f"<|im_start|>user\n{text.strip()}<|im_end|>\n"
        f"<|im_start|>assistant\n"
    )
    ids = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
    if tok.bos_token_id is not None:  # training prepends BOS
        bos = torch.tensor([[tok.bos_token_id]], device=model.device)
        ids["input_ids"] = torch.cat([bos, ids["input_ids"]], dim=1)
        ids["attention_mask"] = torch.cat([torch.ones_like(bos), ids["attention_mask"]], dim=1)
    out = model.generate(**ids, max_new_tokens=256, do_sample=False,
                         eos_token_id=tok.eos_token_id, pad_token_id=tok.pad_token_id)
    return tok.decode(out[0, ids["input_ids"].size(1):], skip_special_tokens=True).strip()

print(translate("لا، لدينا تذاكر يا حبيبتي.", SYS_TO_DAR))

Training

• Base model: oddadmix/Emhotob-5M (Llama arch, hidden 128, 5 layers, 4 heads, vocab 32000, tied embeddings; 5,080,704 params after resizing for 2 ChatML tokens)
• Dataset: oddadmix/darija_english_msa_parallel_dataset (90,104 rows; this model uses the darija and msa columns; rows with an empty msa are skipped). Sources: DoDA, HANTIFARAH combined, and ArabML/Skiredj parallel data.
• Method: HuggingFace Trainer, ChatML, prompt-masked cross-entropy (loss only on the assistant turn). Each row is exploded into two training examples (one per direction). Two ChatML special tokens (<|im_start|>, <|im_end|>) were added and embeddings resized.
• Hyperparameters: 3 epochs · effective batch 64 · LR 3e-4 (cosine, 5% warmup) · bf16 · max length 1024 · load_best_model_at_end on eval_loss.
• Split: 87,104 train / 3,000 deterministic held-out (seed=42), scored both directions.

Limitations

• A 5M model near the emergence threshold: correct register/script but frequent meaning loss, drift, and repetition on longer inputs.
• Darija has no standard orthography and mixes Arabic and Latin (Arabizi) script plus French/Amazigh loanwords — outputs vary in spelling and may hallucinate on out-of-domain input.
• For fluent translation use oddadmix/50M-Darija-MSA-v1.

License

Apache-2.0 (model weights, inherited from the base model). The training dataset aggregates several community corpora — check their individual licenses for downstream use.