OpenSakura/OpenSakura-MODEL-260508-LN-ja-zh-SFT-qwen36-35b-a3b-sft-SHIT

OpenSakura SFT qwen36-35b-a3b checkpoint-5000

WARNING: do not use this model

This model was trained with a garbage dataset that can drop sentences during translation. Do not use this model for production, evaluation, release, or any downstream work until the OpenSakura team has properly organized and retried the dataset generation process and retrained the model.

Known issue: translations may omit input sentences even though the training prompt explicitly asks for one output line per input line.

Provenance

• —Exported checkpoint: /tmp/scratch-space/runs/260513/outputs/checkpoint-5000
• —Base model: OpenSakura/OpenSakura-MODEL-260419-LN-ja-zh-PT-qwen36-35b-a3b-cpt
• —Base revision: 543c3f95d2c08b43e3afcd1232e9b2859c0b264b
• —Recipe: main_lilith65_eve35
• —Training mix: Lilith 0.65, Eve 0.35
• —Lilith format: native_qwen_thinking
• —Eve format: native_qwen_empty_thinking
• —Output dtype: bfloat16

Prompt format used for training

The formatter builds a normal Qwen chat conversation with a system message and a user message, then starts the assistant with Qwen native thinking tags.

System prompt

你是专业的日文轻小说中文译者。请将输入的日文原文翻译为流畅、自然、具有轻小说/二次元语感的简体中文。

要求：
1. 严格保持行数一一对应：输入多少行，输出多少行，不合并、不拆分、不遗漏。
2. 根据上下文自然补全日文省略的主语，并保持角色口吻、性格、性别感和对话关系。
3. 保留原文的特殊格式，包括行首全角空格、转义符、特殊符号等。
4. Ruby/注音格式如“漢字（よみ）”若承载双关、设定、特殊称呼或魔法名，应在译文中保留“汉字（注音）”结构；普通读音标注则直接翻译含义。
5. 日文标点转换为中文标点，如「」转为“”，、转为，。
6. 忠实翻译原文的语气、粗俗度、暴力感、暧昧感和文学风格，不擅自删减、净化、美化或解释。

只输出译文纯文本，不输出说明、分析、JSON、开场白或结束语。

User prompt without glossary

输入(总行数: {line_count})：
{source_text}

User prompt with glossary

When glossary entries are present, the formatter prepends them before the input block:

术语表 <术语原文> -> <术语译文> #<术语信息>:
{source_term_1} -> {target_term_1} #{info_1}
{source_term_2} -> {target_term_2} #{info_2}
输入(总行数: {line_count})：
{source_text}

Chat-template shape

The tokenizer chat template renders the prompt as Qwen chat markers:

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_prompt}<|im_end|>
<|im_start|>assistant
<think>

For Eve empty-thinking examples, the assistant prefill is:

<think>

</think>

{final_translation}<|im_end|>

For Lilith thinking examples, the assistant target is:

<think>
{reasoning_content}
</think>

{final_translation}<|im_end|>

Minimal prompt construction example

This illustrates the training prompt shape only. The model is not recommended for use.

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {
        "role": "user",
        "content": "输入(总行数: 1)：\nこれは例ではありません。",
    },
]

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

Note: the training user prompt header is Chinese 输入(总行数: N)：.

Status

This checkpoint is published only as an internal artifact. It should be replaced after the dataset-generation pipeline is fixed, the data is regenerated, and the model is retrained.