oddadmix

Emhotob-5M-Egyptian-English-v2

README

License: apache-2.0

Evaluation

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

Table with columns: Direction, sacreBLEU, chrF
Direction	sacreBLEU	chrF
English → Egyptian	5.05	25.62
Egyptian → English	10.09	27.37

Saved weights are the best checkpoint by validation loss (eval_loss = 2.467). 20 samples per direction with references are in eval_bidirectional.json.

Example translations

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

English → Egyptian

Table with columns: Source, Model output, Reference
Source	Model output	Reference
So what should I get you? Should I get you this or that? Just imagine the scene: sitting now, wearing a sleeveless under	يعني إيه اللي أقدر أجي لك؟ أنا هاخدك ده ولا دي؟ حتى لو اتنين، اتنين، شوفتوا شوية شوية من الفلوس، كل واحد هيبقى فيه حاجات	طب أجيب لك إيه؟ أجيب لك دي ولا دي؟ تخيل كده المنظر، قاعد دلوقتي لابس فانلة بحمالات وبتاكل أحلى حتة لحمة ممكن تكلها في حي
Prayer niches (mihrabs)	مفيش أي حاجة	محاريب الصلاة

Egyptian → English

Table with columns: Source, Model output, Reference
Source	Model output	Reference
طب أجيب لك إيه؟ أجيب لك دي ولا دي؟ تخيل كده المنظر، قاعد دلوقتي لابس فانلة بحمالات وبتاكل أحلى حتة لحمة ممكن تكلها في حي	Well, what do you get to you? So if you get to you, it's not like this, it's just what's going on inside, she's just wai	So what should I get you? Should I get you this or that? Just imagine the scene: sitting now, wearing a sleeveless under
محاريب الصلاة	My dear, the way I'm going to see.	Prayer niches (mihrabs)

Usage

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

python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "oddadmix/Emhotob-5M-Egyptian-English-v2"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, dtype=torch.bfloat16).to("cuda").eval()

SYSTEM = "أنت مترجم محترف. ترجم النص الإنجليزي إلى اللهجة المصرية العامية."

def translate(text, system=SYSTEM):
    prompt = (f"<|im_start|>system\n{system}<|im_end|>\n"
              f"<|im_start|>user\n{text.strip()}<|im_end|>\n<|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:
        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()

Training

Base model: oddadmix/Emhotob-5M-v2 (Llama arch, hidden 128, 5 layers, 4 heads, vocab 32000, tied embeddings; 5,080,704 params after resizing for 2 ChatML tokens)
Dataset: oddadmix/egyptian-translation-dataset-2.9-openai-batch
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).
Hyperparameters: 3 epochs · effective batch 64 · LR 3e-4 (cosine, 5% warmup) · bf16 · max length 1024 · load_best_model_at_end on eval_loss.
Eval split: 3,000 deterministic held-out pairs (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.

Table with columns: Metric, Score
Metric	Score
BLEU (`evaluate`, 0–1 — leaderboard metric)	0.0108
BLEU (sacrebleu, 0–100)	1.08
chrF	22.19
METEOR	0.1352

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

Table with columns: Model, Params, BLEU (hf), BLEU (sacre), chrF, METEOR
Model	Params	BLEU (hf)	BLEU (sacre)	chrF	METEOR
5M v1	5.2M	0.0000	0.19	13.00	0.0558
5M v2 (this model)	5.2M	0.0108	1.08	22.19	0.1352
10M v1	11.2M

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 ~5.1M model: reliable on short/common sentences, but drift, repetition, and errors appear on long or rare inputs. Gender is disambiguated only from context. For fluent translation use the 50M sibling.

License

Apache-2.0, inherited from the base model.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

oddadmix

Model Tree

Base

oddadmix/Emhotob-5M-v2

Fine-tuned

this model

Input Modalities

Text

Output Modalities