Respair/Hayate_Translate_FT_EN2JP

Inference

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "Respair/Hayate_Translate_FT_EN2JP"
tokenizer = "Qwen/Qwen3-4B-Instruct-2507"

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

text = (

    "Mitochondria produce ATP by coupling electron transport to proton translocation across the inner membrane. "
    "As electrons move through the respiratory chain, protons are pumped into the intermembrane space, and ATP synthase harnesses their return flow to phosphorylate ADP."
    "The coupling is imperfect: some protons leak back independently, dissipating the gradient as heat."

).replace('\n', ' ').strip()


prompt = "JP->EN: "

messages = [
    {"role": "user", "content": prompt + text}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=8192,
    do_sample=True,
    temperature=0.1, # or whatever
    top_p=0.95,
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist() 

content = tokenizer.decode(output_ids, skip_special_tokens=True)

print(content)

# out:
# ミトコンドリアは、内膜を挟んだプロトン輸送と電子輸送をカップリングしてATPを産生する。
# 呼吸鎖を電子が通る過程で、プロトンが膜間腔に汲み出され、その戻り流しを利用してATP合成酵素がADPをリン酸化する。
# カップリングは不完全で、一部のプロトンは独立して戻り、勾配を熱として散逸する。

Details

This is a translation model for English -> Japanese.
for Japanese to English please visit Respair/Hayate_Translate_base_v1.0_JPtoEN I trained it mainly because I needed something lightweight with a long enough context window for a few personal projects.

these models were trained on roughly 4-7B tokens of high quality data, with a context length of 10k tokens, using a few tricks to source and clean the data.

Japanese is highly contextual in ways English doesn't always encode, but I tried to cover as many topics as possible, from fiction to science and code. it should fare quite well given its smaller size. It is also programming syntax-aware, and will sometimes automatically provide the original technical phrase in parentheses.

Benchmarking

The benchmarking method is a simple LLM as a Judge on ~2400 samples (you can find it in this repo), covering multiple topics.
I used Gemini 3 Pro to judge the naturalness, accuracy and overall flow of the output, comparing it with GPT 5.2 (no reasoning) & CAT Translate.

Gemini prefers this model in the majority of cases.

Only English -> Japanese was assessed.

Note:

• Translation is a one-to-many problem, so this benchmarking method doesn't necessarily determine which model is more accurate.