RFMT at WMT 2026 General Translation Task

This is Team RFMT's submission system for the WMT2026 constrained (open-weight) track of the General Machine Translation Task. It supports English->Japanese and Chinese->Japanese translation. In addition to standard translation, it supports:

• FIM (Fill-in-the-Middle) translation
• Specification of domain, style, and glossary

Usage (Transformers)

Verified to work with version 5.5.4.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tohoku-nlp/RFMT"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto"
)

text = "Today felt simple but warm, with a quiet morning and a few small moments that made me smile. I took a walk, noticed the soft light in the sky, and felt grateful for the peaceful rhythm of the day. By evening, I realized that even an ordinary day can become a gentle memory when I pay attention to it."

message = [
    {"role": "user", "content": text}
]
glossary = [
    {"source": "quiet morning", "target": "静寂な朝"},
    {"source": "ordinary day", "target": "凡庸な日"},
]


inputs = tokenizer.apply_chat_template(
    message,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    domain="News",
    style="だ・である調",
    glossary=glossary,
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(
    tokenizer.decode(
        outputs[0][inputs["input_ids"].shape[-1]:], 
        skip_special_tokens=True
    )
)

Usage (vLLM)

Verified to work with version 0.19.1.

Launching the server

vllm serve tohoku-nlp/RFMT \
    --gpu-memory-utilization 0.95 \
    --dtype bfloat16 \
    --seed 42

Sending requests

PORT=8000
# General translation
curl http://localhost:$PORT/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "X-Request-Id: general_translation_example" \
  -d '{
    "model": "tohoku-nlp/RFMT",
    "messages": [
      {
        "role": "user",
    "content": "Today felt simple but warm, with a quiet morning and a few small moments that made me smile. I took a walk, noticed the soft light in the sky, and felt grateful for the peaceful rhythm of the day. By evening, I realized that even an ordinary day can become a gentle memory when I pay attention to it."
      }
    ],
    "max_tokens": 512,
    "temperature": 0.0
  }' | jq


# FIM Translation
curl http://localhost:$PORT/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "X-Request-Id: fim_translation_example" \
  -d '{
    "model": "tohoku-nlp/RFMT",
    "messages": [
      {
        "role": "user",
    "content": "Today felt simple but warm, with a quiet morning and a few small moments that made me smile. <|fim_prefix|>I took a walk, noticed the soft light in the sky, and felt grateful for the peaceful rhythm of the day. <|fim_suffix|>By evening, I realized that even an ordinary day can become a gentle memory when I pay attention to it."
      },
      {
    "role": "assistant",
    "content": "今日はシンプルだけど温かい一日で、静かな朝と、私を笑顔にしてくれる小さな瞬間がいくつかありました。<|fim_middle|>夕方になると、普通の日でも注意深く見ると優しい思い出になることに気づきました。"
      }
    ],
    "max_tokens": 512,
    "temperature": 0.0,
    "chat_template_kwargs": {"add_fim_generation_prompt": true}
  }' | jq


# Context-Aware Incremental Translation
curl http://localhost:$PORT/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "X-Request-Id: prefix_forcing_example" \
  -d '{
    "model": "tohoku-nlp/RFMT",
    "messages": [
      {
        "role": "user",
    "content": "Today felt simple but warm, with a quiet morning and a few small moments that made me smile. I took a walk, noticed the soft light in the sky, and felt grateful for the peaceful rhythm of the day. By evening, I realized that even an ordinary day can become a gentle memory when I pay attention to it."
      }
    ],
    "max_tokens": 512,
    "temperature": 0.0,
    "chat_template_kwargs": {"generation_prefix": "今日はとりとめもなく暖かい"}
  }' | jq

# Domain prompting
curl http://localhost:$PORT/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "X-Request-Id: domain_specification_example" \
  -d '{
    "model": "tohoku-nlp/RFMT",
    "messages": [
      {
        "role": "user",
    "content": "Today felt simple but warm, with a quiet morning and a few small moments that made me smile. I took a walk, noticed the soft light in the sky, and felt grateful for the peaceful rhythm of the day. By evening, I realized that even an ordinary day can become a gentle memory when I pay attention to it."
      }
    ],
    "max_tokens": 512,
    "temperature": 0.0,
    "chat_template_kwargs": {"domain": "News"}
  }' | jq


# Style prompting
curl http://localhost:$PORT/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "X-Request-Id: style_specification_example" \
  -d '{
    "model": "tohoku-nlp/RFMT",
    "messages": [
      {
        "role": "user",
    "content": "Today felt simple but warm, with a quiet morning and a few small moments that made me smile. I took a walk, noticed the soft light in the sky, and felt grateful for the peaceful rhythm of the day. By evening, I realized that even an ordinary day can become a gentle memory when I pay attention to it."
      }
    ],
    "max_tokens": 512,
    "temperature": 0.0,
    "chat_template_kwargs": {"style": "だ・である調"}
  }' | jq



# Glossary prompting
curl http://localhost:$PORT/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "X-Request-Id: glossary_specification_example" \
  -d '{
    "model": "tohoku-nlp/RFMT",
    "messages": [
      {
        "role": "user",
    "content": "Today felt simple but warm, with a quiet morning and a few small moments that made me smile. I took a walk, noticed the soft light in the sky, and felt grateful for the peaceful rhythm of the day. By evening, I realized that even an ordinary day can become a gentle memory when I pay attention to it."
      }
    ],
    "max_tokens": 512,
    "temperature": 0.0,
    "chat_template_kwargs": {"glossary": [{"source": "quiet morning", "target": "静寂な朝"}, {"source": "ordinary day", "target": "凡庸な日"}]}
  }' | jq

Disclaimer

While the authors of this model/dataset have exercised utmost care regarding its contents, functionality, and other aspects during its creation, they make no guarantees as to the accuracy or safety of the model's outputs and assume no responsibility whatsoever.
In the unlikely event that any inconvenience or damage occurs to the user as a result of using this model/dataset, neither the authors of the model or dataset nor the organizations to which the authors belong shall bear any responsibility.

License

This model and dataset are distributed under the Apache License 2.0.

Acknowledgments

We would like to thank everyone at the Tohoku NLP Group for their cooperation in various aspects of building this model. We would also like to express our gratitude to everyone who organized the WMT General MT Task.

Team Members

• Ryosuke Matsuda (Tohoku University)
• Keito Kudo (Tohoku University / RIKEN)
• Ryo Fujii (Future Corporation / Tohoku University)
• Takumi Ito (Tohoku University / Machine Learning Solutions)
• Makoto Morishita (Future Corporation / Tohoku University)
• Jun Suzuki (Tohoku University / RIKEN)

Citation

Work in progress. We are currently writing a system description paper.