50M-Egyptian-English-v1 — Bidirectional English ↔ Egyptian Arabic

A 51.8M-parameter small language model that translates both ways between English and Egyptian colloquial Arabic (المصرية العامية). A single set of weights serves both directions; a direction-specific system prompt selects which way to translate.

Finetuned from oddadmix/50M-2048-Emhotob, a tiny Arabic base model trained from scratch.

Evaluation

Evaluated on a deterministic held-out set of 3,000 pairs (seed=42), decoded greedily (do_sample=False, no repetition penalty), scored with sacreBLEU:

Direction	sacreBLEU	chrF
English → Egyptian	23.83	51.94
Egyptian → English	33.90	53.01

Egyptian→English scores notably higher, as expected — English is more standardized so a single reference captures more of the valid output space, whereas Egyptian colloquial has many equally-correct spellings and phrasings. The saved weights are the best checkpoint by validation loss (eval_loss=0.9297, epoch 2 of 3).

Decoding note: use plain greedy. A repetition penalty (1.2) was tested on these 50M translation models and lowered BLEU by 5–12 points, because Arabic legitimately repeats short particles that the penalty suppresses.

Example translations

Real greedy-decoded outputs from the held-out set:

English → Egyptian

English input	Model output (Egyptian)
Man, these things happen. Sometimes Liverpool loses the match, and other times Real Madrid wins it.	يا عم الحاجات دي بتحصل. ساعات ليفربول بيخسر الماتش، وساعات ريال مدريد بيكسبه.
I will be fine, don't worry	أنا هبقى كويس، ما تقلقش
I want grilled chicken, kofta on charcoal, and Pepsi.	عايز فراخ مشوية وكفتة على الفحم وبابسي.

Egyptian → English

Egyptian input	Model output (English)
يا عم الحاجات دي بتحصل، ساعات ليفربول بتخسر الماتش وساعات ريال مدريد بيكسب.	Hey, these things happen, sometimes Liverpool loses the match and sometimes Real Madrid wins.
انا هبقى كويسة، ما تقلقيش	I'll be fine, don't worry
انا عايز فراخ مشوية وكفتة على الفحم وبيبسي.	I want grilled chicken, kofta, and chopped on charcoal.

A larger set of 20 examples per direction (with references) is 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/50M-Egyptian-English-v1"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16).to("cuda").eval()

SYS_TO_EGY = "أنت مترجم محترف. ترجم النص الإنجليزي إلى اللهجة المصرية العامية."
SYS_TO_EN  = "You are a professional translator. Translate the Egyptian Arabic text into English."

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("I will be fine, don't worry", SYS_TO_EGY))
# → أنا هبقى كويس، ما تقلقش
print(translate("انا هبقى كويسة، ما تقلقيش", SYS_TO_EN))
# → I'll be fine, don't worry

Training

• Base model: oddadmix/50M-2048-Emhotob (Llama arch, ~51.8M params)
• Dataset: oddadmix/egyptian-translation-dataset-2.9-openai-batch (135K rows, English/Arabic columns; Arabic = Egyptian colloquial)
• 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, ~261K total). 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: 132,282 train / 3,000 deterministic held-out (seed=42), scored both directions.

Out-of-domain evaluation — Egyptian Arabic Translation Benchmark

The results above are in-domain: a held-out split of the same corpus this model was trained on. The numbers below are out-of-domain — the same model scored on the Egyptian Arabic Translation Benchmark (oddadmix/egyptian-arabic-translation-benchmark, 319 English→Egyptian pairs written by a different annotator with different orthographic conventions).

Expect these to be substantially lower than the in-domain scores. That gap is the generalization penalty, not a regression — both numbers are real, they measure different things.

Metric	Score
BLEU (evaluate, 0–1 — leaderboard metric)	0.1113
BLEU (sacrebleu, 0–100)	11.13
chrF	44.41
METEOR	0.3482

Decoding is deterministic greedy (do_sample=False, no repetition penalty), using the exact ChatML prompt format the model was trained with — the same protocol as every other number in this study.

Where this rung sits

Model	Params	BLEU (hf)	BLEU (sacre)	chrF	METEOR
5M v1	5.2M	0.0000	0.19	13.00	0.0558
5M v2	5.2M	0.0108	1.08	22.19	0.1352
10M v1	11.2M	0.0313	3.13	28.52	0.1985
10M v2	10.9M	0.0341	3.41	31.60	0.2235
25M v1	25.3M	0.0991	9.91	41.07	0.3223
25M v2	25.3M	0.0824	8.24	40.29	0.3148
50M v1 (bidi) (this model)	51.8M	0.1113	11.13	44.41	0.3482
50M v1 (uni)	51.8M	0.1261	12.61	46.13	0.3559

Reading these numbers

BLEU understates quality on this set. Scoring is against a single reference, so a correct translation that picks a different valid word is penalized — e.g. فريش vs the reference's طازة for "fresh", or التليفون اللي ضاع vs تليفونها الضايع for "her lost phone". Both are good Egyptian; only one matches the reference. chrF and METEOR track perceived quality more closely here.

At 319 rows, differences of roughly 1–2 BLEU between adjacent rungs are within noise.

These are small models — 5M to 50M parameters, orders of magnitude below the large systems typically evaluated on this benchmark. The result of interest is the scaling curve and per-parameter efficiency, not absolute rank against models 100–1000× the size.

Limitations

• A 50M model: expect errors on rare / technical vocabulary, proper nouns, and occasional drift on long inputs (see the "Haman"/"chopped" slips in the examples). Everyday conversational text is handled well.
• Gender is disambiguated only from context; ambiguous inputs may default one way.
• Trained on conversational Egyptian ↔ English; other Arabic dialects and formal/ technical registers are out of scope.

License

Apache-2.0, inherited from the base model.