Qwen3_LIL_en2zh_translation

Qwen3_LIL_en2zh_translation

Model Details

利用带有噪声的粗糙翻译数据，通过大模型的推理能力对翻译结果进行修正，并在链式思维（CoT）中引入目标句的上下文信息。

最终将修正后的翻译与对应的思维链用于小模型训练，在实现数据净化的同时，保留了原始数据中蕴含的知识与大模型的思考能力。

By leveraging a noisy, imperfect translation dataset, the model uses the reasoning capabilities of a large-scale model to refine translations. Context around the translation target is incorporated through chain-of-thought (CoT) prompting.

The resulting CoT traces and corrected translations are then used to train a smaller model, achieving data purification while preserving both the implicit knowledge in the original dataset and the reasoning ability of the thinking model.

Model Description

• Model type: Lora adapters
• Language(s) (NLP): English to Chinese
• License: apache-2.0
• Finetuned from model [optional]: Qwen3-1.7B Qwen3-4B

Uses

注意：本模型在 BnB 量化下进行了 LoRA 微调，

由于微调权重无法直接合并到主模型权重中，因此仅提供 LoRA adapter 文件。

Note: This model was fine-tuned using LoRA under BnB quantization.

Since the fine-tuned weights cannot be directly merged into the base model, only the LoRA adapter files are provided.

思考模式 / Thinking mod

# in_file_name 所在文件名
# pre_text，next_text <|speaker_name_A|>dialogue_1<|speaker_name_B|>dialogue_2.....
# target_text <|speaker_name_TAR|>dialogue_TAR

prompt = f"<|im_start|>user\n文件：{in_file_name}\n上下文：{pre_text}{target_text}{next_text}\n目标原文：{target_text}<|im_end|>\n<|im_start|>assistant\n"

model_name = "../dl_models/Qwen3-4B"
lora_model_name = "Qwen3-4B-LIL-DIS-LoRA-16-30000"
device = "cuda" if torch.cuda.is_available() else "cpu"

Q_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

# load the tokenizer and the model
tokenizer: Qwen2TokenizerFast = AutoTokenizer.from_pretrained(model_name)
model: Qwen3ForCausalLM = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=Q_config,
    # attn_implementation="flash_attention_2"
)

lora_model = PeftModel.from_pretrained(model, lora_model_name).to(device)

model_input_batch = tokenizer([prompt],return_tensors="pt").to(device)
output_ids = lora_model.generate(**model_input_batch,max_new_tokens=2048)

res = tokenizer.batch_decode(output_ids[:, model_input_batch.input_ids.shape[1]:],skip_special_tokens=True,)
print(res[0])

直接推理模式 / NO Thinking mod

# in_file_name 所在文件名
# pre_text，next_text <|speaker_name_A|>dialogue_1<|speaker_name_B|>dialogue_2.....
# target_text <|speaker_name_TAR|>dialogue_TAR

prompt = f"文件：{in_file_name}\n上下文：{pre_text}{target_text}{next_text}\n目标原文：{target_text}\n翻译："

model_name = "../dl_models/Qwen3-1.7B"
lora_model_name = "Qwen3-1.7B-bnb4-LIL-16-30000"
device = "cuda" if torch.cuda.is_available() else "cpu"

Q_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

# load the tokenizer and the model
tokenizer: Qwen2TokenizerFast = AutoTokenizer.from_pretrained(model_name)
model: Qwen3ForCausalLM = AutoModelForCausalLM.from_pretrained(
    model_name,
    quantization_config=Q_config,
    # attn_implementation="flash_attention_2"
)

lora_model = PeftModel.from_pretrained(model, lora_model_name).to(device)

model_input_batch = tokenizer([prompt],return_tensors="pt").to(device)
output_ids = lora_model.generate(**model_input_batch,max_new_tokens=512)

res = tokenizer.batch_decode(output_ids[:, model_input_batch.input_ids.shape[1]:],skip_special_tokens=True,)
print(res[0])

Limitations

本微调模型是在小规模的领域专有数据集上训练的，无法保证在其他领域的表现。

加载模型时请勿使用 merge_and_unload()，否则可能导致失败。

This fine-tuned model was trained on a small, domain-specific private dataset and may not generalize well to other domains.

Do not use merge_and_unload() when loading the model, as it may cause failure.

Training Details

本项目使用 Qwen-14B-AWQ 作为教师模型，启用思考模式，并设置了合适的 system prompt，对粗糙的翻译进行润色与规范统一，生成了 30,000 条带有链式思维（CoT）的翻译结果。

随后在 Qwen-4B 上进行 QLoRA 微调，保留 CoT；在 Qwen-1.7B 上进行 QLoRA 微调，但不使用 CoT。

Qwen-14B-AWQ was used as the teacher model, with reasoning mode enabled and an appropriate system prompt configured to polish and standardize noisy translations. This process generated 30,000 translation samples with chain-of-thought (CoT) traces.

QLoRA fine-tuning was then performed on Qwen-4B with CoT retained, and on Qwen-1.7B without using CoT.

Training Data

私有数据 / Private data

Training Hyperparameters

Q_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.bfloat16,
)

L_config = LoraConfig(
    r=lora_rank,
    use_rslora=True,
    target_modules=["q_proj", "v_proj", "k_proj"],
    lora_dropout=0.1,
    bias="none",
)

optimizer = AdamW8bit(lora_model.parameters(), lr=1e-5, betas=(0.9, 0.95))

step_num = 30000
scheduler = get_wsd_schedule(
    optimizer=optimizer,
    num_warmup_steps=step_num * 0.05,
    num_stable_steps=step_num * 0.85,
    num_decay_steps=step_num * 0.1 + 1
)

batch_size = 8

Model Examination

一方面，如何在微调过程中使用普通数据集，同时保留模型原有的思考能力（CoT），一直是困扰部分研究者的问题；

另一方面，小规模领域数据集通常质量较低、噪声较多，如何高效利用也是一个挑战。

因此，我尝试探索一种方法：利用大模型在优化翻译时所产生的思考过程，来训练小模型，并取得了一定的效果。

不过，这种方法是否具有通用性、是否会影响模型的原始能力，以及是否存在不依赖教师模型的替代方案，目前仍不明确。

On one hand, how to fine-tune a model using general-purpose datasets while preserving its chain-of-thought (CoT) reasoning remains a challenge for some researchers.

On the other hand, small domain-specific datasets often suffer from low quality and high noise, making effective utilization difficult.

To address this, I explored a method that trains smaller models using the reasoning traces generated by large models during translation refinement, and this has shown some promising results.

However, it is still unclear whether this method is broadly applicable, whether it affects the model’s original capabilities, or whether alternatives that do not rely on teacher models exist.