YehaTranslate

YehaTranslate is a fine-tuned translation model for Ethiopian languages ↔ English, developed by Hasab AI. It is based on google/translategemma-4b-it and fine-tuned on paired Amharic–English data.

The primary training focus is Amharic (am) ↔ English (en). The training data also contains samples in Tigrinya (ti) and Oromo (om), so translation to/from English is supported for these languages — though coverage is not as extensive as Amharic. We plan to expand Tigrinya and Oromo training data in future releases.

Model Details

Field	Value
Base model	google/translategemma-4b-it
Architecture	Gemma3ForConditionalGeneration (4B)
Languages	Amharic (am), Tigrinya (ti), Oromo (om), English (en)
Framework	PyTorch + HuggingFace Transformers, distributed with torchrun

Benchmark Results

Evaluated on 960 held-out sentence pairs per direction. Primary metric: chrF2 (character-level F-score, β=2), which is more reliable than BLEU for morphologically rich Amharic.

am → en

Model	BLEU	chrF2	Len Ratio
Baseline Gemma (no fine-tuning)	7.00	27.69	1.22
YehaTranslate (ours)	37.58	64.60	1.07
Gemini 2.5 Flash	36.17	57.78	0.69
GPT-5.4	26.37	51.36	0.67

en → am

Model	BLEU	chrF2	Len Ratio
Baseline Gemma (no fine-tuning)	1.00	7.69	1.79
YehaTranslate (ours)	8.12	25.93	1.13
Gemini 2.5 Flash	4.43	23.05	0.43
GPT-5.4	0.15	9.51	0.18

Usage

import torch
from transformers import AutoProcessor, AutoModelForImageTextToText

model_id = "hasab-ai/YehaTranslate"

processor = AutoProcessor.from_pretrained(model_id, use_fast=True)
model = AutoModelForImageTextToText.from_pretrained(model_id, torch_dtype=torch.bfloat16)
model.generation_config.pad_token_id = processor.tokenizer.eos_token_id
model.eval()

DIRECTIONS = {
    "Amharic → English":  ("am", "en"),
    "English → Amharic":  ("en", "am"),
    "Oromo → English":    ("om", "en"),
    "English → Oromo":    ("en", "om"),
    "Tigrinya → English": ("ti", "en"),
    "English → Tigrinya": ("en", "ti"),
}

def translate(text: str, src_lang: str, tgt_lang: str, max_new_tokens: int = 512) -> str:
    messages = [{"role": "user", "content": [
        {"type": "text", "source_lang_code": src_lang,
         "target_lang_code": tgt_lang, "text": text.strip()}
    ]}]
    inputs = processor.apply_chat_template(
        messages,
        tokenize=True,
        add_generation_prompt=True,
        return_tensors="pt",
        return_dict=True,
    )
    input_len = inputs["input_ids"].shape[1]
    with torch.inference_mode():
        out = model.generate(**inputs, do_sample=False, max_new_tokens=max_new_tokens)
    return processor.decode(out[0][input_len:], skip_special_tokens=True).strip()

# Amharic → English
print(translate("የኢትዮጵያ ሕዝቦች ታሪካዊ ትብብር እና አንድነት ዛሬም ቀጥሏል።", "am", "en"))

# English → Amharic
print(translate("The coffee ceremony is an important part of Ethiopian culture.", "en", "am"))

# Oromo → English
print(translate("Biyyoonni Itoophiyaa beekamtii addunyaa qabdi.", "om", "en"))

Training Details

Training used torchrun for distributed data-parallel training, with the HuggingFace Trainer API and bfloat16 activations.

The model was trained on paired Amharic–English sentence pairs covering news, general web, and social media domains.

Limitations

• Best performance on general-domain text; domain-specific content (legal, medical) may degrade.
• en→am generation is harder than am→en decoding, consistent with the asymmetry seen across all models.
• Amharic is morphologically rich; single-reference automatic metrics (BLEU, chrF2) underestimate translation quality when paraphrase variation is high.

Citation

@misc{oriontranslate2026,
  title        = {YehaTranslate: Fine-tuned Amharic-English Translation},
  author       = {Hasab AI},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/hasab-ai/YehaTranslate}},
}