Model Card for Llama-2-7B-EAX

Model Details

Model Description

Llama-2-7B-EAX is a language model specifically enhanced for inter non-English language pairs. The model is built on top of Llama 2, following a two-stage training approach: first, an English-centric parallel data supervised fine-tuning stage (the SFT model is available at Llama-2-7b-MT-SFT), followed by a dedicated x2x optimization stage. This approach strategically leverages the established English-centric capabilities of large language models to bootstrap comprehensive multilingual translation capabilities.

• Model type: A 7B parameter translation model built on top of Llama 2, enhanced for x2x language pairs through specialized optimization.
• Language(s) (NLP): English, Portuguese, Spanish, French, German, Dutch, Italian, Korean, Russian, Chinese
• License: The LLAMA 2 Community License, Copyright © Meta Platforms, Inc. All Rights Reserved.

Intended uses & limitations

Llama-2-7B-EAX is designed for direct translation between non-English language pairs, addressing a significant gap in current LLM translation capabilities. The model maintains strong performance on English-centric translation while significantly improving x2x translation quality.

Here's how you can run the model with Huggingface Transformers:

from transformers import AutoTokenizer, AutoModelForCausalLM

MODEL_PATH = "double7/Llama-2-7b-EAX"

tokenizer = AutoTokenizer.from_pretrained(MODEL_PATH)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_PATH, device_map="auto", torch_dtype="auto"
)

src_lang = "German"
trg_lang = "Chinese"
src_text = "Filmkarriere Collinges Filmdebüt in Die kleinen Füchse von 1941 brachte ihr eine Nominierung für den Academy Award als beste Nebendarstellerin ein."

prompt = f"Translate the following text from {src_lang} into {trg_lang}:\n{src_lang}: {src_text}\n{trg_lang}:"

# We use the tokenizer’s chat template to format each message - see https://huggingface.co/docs/transformers/main/en/chat_templating
messages = [
    {"role": "user", "content": prompt},
]

input_text = tokenizer.apply_chat_template(
    messages, tokenize=False, add_generation_prompt=True
)

inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, do_sample=False, max_new_tokens=256)
output_text = tokenizer.batch_decode(outputs, skip_special_tokens=False)[0]
print(output_text)
# <s><|im_start|> user
# Translate the following text from German into Chinese:
# German: Filmkarriere Collinges Filmdebüt in Die kleinen Füchse von 1941 brachte ihr eine Nominierung für den Academy Award als beste Nebendarstellerin ein.
# Chinese:<|im_end|> 
# <|im_start|> assistant
#  科林格的电影生涯开始于 1941 年的《小狐狸》，她因此获得了奥斯卡最佳女配角提名。<|im_end|>

Translation Instructions

Following TowerInstruct, we use diverse translation instructions in training, you can use natural language to describe translation requests, such as:

prompt1 = f"Translate the following text from {src_lang} into {trg_lang}:\n{src_lang}: {src_text}\n{trg_lang}:"

prompt1 = f"Please provide a translation from {src_lang} to {trg_lang} for the following text:\n{src_text}\nTarget:",

prompt2 = f"Translate this {src_lang} text into {trg_lang}:\nSource: {src_text}\nTranslation:",

We use prompt1 for the evaluation.

Out-of-Scope Use

The model is not guaranteed to perform for languages other than the 10 languages it supports.

Bias, Risks, and Limitations

Llama-2-7B-EAX has not been aligned to human preferences, so the model may generate problematic outputs (e.g., hallucinations, harmful content, or false statements).

Prompt Format

Llama-2-7B-EAX was trained using the ChatML prompt templates without any system prompts. An example follows below:

<|im_start|>user
{USER PROMPT}<|im_end|>
<|im_start|>assistant
{MODEL RESPONSE}<|im_end|>
<|im_start|>user
[...]

Training Details

Training Data

We use ~140k high-confidence synthetic data for optimization. This data is based on Llama-2-7b and the translation data from TowerBlocks as seeds, and was curated through our specialized pipeline. See our paper for more details.

Training hyperparameters

The following hyperparameters were used during x2x training:

• learning_rate: 2e-07
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1
• max_seq_length: 2048
• DPO beta: 0.2
• SFT coefficient: 2.0

Citation

@misc{yang2025enanchoredx2xenglishanchoredoptimizationmanytomany,
      title={EnAnchored-X2X: English-Anchored Optimization for Many-to-Many Translation}, 
      author={Sen Yang and Yu Bao and Yu Lu and Jiajun Chen and Shujian Huang and Shanbo Cheng},
      year={2025},
      eprint={2509.19770},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2509.19770}, 
}