Tower-7B-Stair_2c1t

This model is released as part of our paper Doc2FRC: Length-Consistent Document-Level Machine Translation via Fixed-Range Chunking. The code and paper-specific inference scripts are available in the Doc2FRC GitHub repository.

Tower-7B-Stair_2c1t is a full-parameter fine-tuned version of Unbabel/TowerInstruct-Mistral-7B-v0.2 for multilingual chunk-level machine translation. It is the Stair_2c1t variant: it uses a stair-step context scheme based on chunk position. The first chunk is translated without preceding context, the second chunk uses one preceding source-language chunk, and later chunks use the two immediately preceding source-language chunks, supplied separately through the Context1 and Context2 sections. The training examples were derived from sardinelab/DocBlocks using fixed-range chunking with a 256–512-token range.

Supported translation directions

The model supports translation between English and the following languages in both directions:

• German
• Spanish
• French
• Italian
• Korean
• Dutch
• Portuguese
• Russian
• Chinese

General usage

The example below demonstrates general model usage for translating a current source chunk with zero, one, or two preceding source-language context chunks. For the exact document chunking, inference scripts, prompting setup, and evaluation procedure used in the paper, please refer to the Doc2FRC GitHub repository.

Recommended prompt formats

The model was fine-tuned with three raw ChatML-style translation prompt formats, selected according to the current chunk position.

First chunk: 0c1t

<|im_start|>user
Translate the following source text from {SOURCE_LANGUAGE} into {TARGET_LANGUAGE}.
{SOURCE_LANGUAGE}: {SOURCE_TEXT}.
{TARGET_LANGUAGE}: <|im_end|>
<|im_start|>assistant

Second chunk: 1c1t

<|im_start|>user
Context
{SOURCE_LANGUAGE}: {CONTEXT_TEXT}
Translate the following source text from {SOURCE_LANGUAGE} into {TARGET_LANGUAGE}.
{SOURCE_LANGUAGE}: {SOURCE_TEXT}.
{TARGET_LANGUAGE}: <|im_end|>
<|im_start|>assistant

Later chunks: 2c1t

<|im_start|>user
Context1
{SOURCE_LANGUAGE}: {CONTEXT_TEXT_1}
Context2
{SOURCE_LANGUAGE}: {CONTEXT_TEXT_2}
Translate the following source text from {SOURCE_LANGUAGE} into {TARGET_LANGUAGE}.
{SOURCE_LANGUAGE}: {SOURCE_TEXT}.
{TARGET_LANGUAGE}: <|im_end|>
<|im_start|>assistant

Use full English language names such as English, Chinese, German, or Russian. For the 2c1t format, CONTEXT_TEXT_1 is the earlier context chunk and CONTEXT_TEXT_2 is the immediately preceding context chunk.

Transformers example

Install a PyTorch build appropriate for your hardware, together with Transformers and Accelerate. PyTorch 2.6 or later is recommended for loading the current PyTorch .bin checkpoint files.

pip install "transformers>=4.56.2" accelerate

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_ID = "ynklab/Tower-7B-Stair_2c1t"

tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID,
    dtype=torch.bfloat16,
    device_map="auto",
)
model.eval()


def build_prompt(
    source_language,
    target_language,
    source_text,
    context_texts=None,
):
    context_texts = context_texts or []
    if len(context_texts) > 2:
        raise ValueError("Stair_2c1t accepts at most two context chunks.")

    lines = ["<|im_start|>user"]

    if len(context_texts) == 1:
        lines.extend([
            "Context",
            f"{source_language}: {context_texts[0]}",
        ])
    elif len(context_texts) == 2:
        lines.extend([
            "Context1",
            f"{source_language}: {context_texts[0]}",
            "Context2",
            f"{source_language}: {context_texts[1]}",
        ])

    lines.extend([
        f"Translate the following source text from {source_language} "
        f"into {target_language}.",
        f"{source_language}: {source_text}.",
        f"{target_language}: <|im_end|>",
        "<|im_start|>assistant",
        "",
    ])
    return "\n".join(lines)


source_language = "English"
target_language = "Chinese"
source_text = "The weather is nice today"

# Use [] for the first chunk, one item for the second chunk,
# and the two immediately preceding source chunks for later chunks.
context_texts = [
    "We planned a picnic for this afternoon.",
    "We checked the forecast before leaving.",
]
prompt = build_prompt(
    source_language,
    target_language,
    source_text,
    context_texts,
)

# The Tower tokenizer adds the beginning-of-sequence token used during training.
inputs = tokenizer(prompt, return_tensors="pt")
inputs = {name: tensor.to(model.device) for name, tensor in inputs.items()}

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=False,
        repetition_penalty=1.05,
    )

generated_tokens = outputs[0, inputs["input_ids"].shape[1]:]
translation = tokenizer.decode(
    generated_tokens,
    skip_special_tokens=True,
).strip()

print(translation)

For paper-level document translation, split the document into fixed-range chunks and reconstruct the translated chunks using the procedures provided in the Doc2FRC repository.

Training

• Base model: TowerInstruct-Mistral-7B-v0.2
• Training method: full-parameter supervised fine-tuning
• Training variant: Stair_2c1t (stair-step mixture of zero, one, or two preceding source-language context chunks and one target chunk)
• Fixed-range segmentation: 256–512 tokens
• Epochs: 2
• Learning rate: 7e-6
• Learning-rate scheduler: cosine
• Warmup steps: 125
• Maximum sequence length: 32,768 tokens
• Training precision: bfloat16
• Optimizer: AdamW
• Weight decay: 0.01

License

This model preserves the CC BY-NC-SA 4.0 License distributed with its base model, TowerInstruct-Mistral-7B-v0.2. See the LICENSE file and the upstream model card for the applicable terms.

DocBlocks contains material derived from multiple sources. Users should also consult the DocBlocks dataset and the original data sources for their applicable licensing conditions.

Acknowledgements

This model is based on TowerInstruct-Mistral-7B-v0.2 and was fine-tuned using DocBlocks. Please cite our paper when using this model in academic work.

Citation

@misc{wang2026doc2frclengthconsistentdocumentlevelmachine,
      title={Doc2FRC: Length-Consistent Document-Level Machine Translation via Fixed-Range Chunking}, 
      author={Xiaotian Wang and Youyuan Lin and Zhan Shen and Hitomi Yanaka},
      year={2026},
      eprint={2609.12674},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2609.12674}, 
}