同样的数据集，Qwen3-0.6B，通过train.py脚本微调出现重复和胡言乱语

[skyler9901]

请问博主有遇到过这样的问题吗？

[skyler9901]

但是在inference.py中对output的checkpoint进行推理，就是正确输出的，这是为什么呀，请问博主知道吗？请教请教，小菜狗真的很想知道~~

[yiiizhang]

但是在inference.py中对output的checkpoint进行推理，就是正确输出的，这是为什么呀，请问博主知道吗？请教请教，小菜狗真的很想知道~~

我也发现了，是某些梯度、mask还是什么导致的呀？

[ppppps]

啊我也遇到了这个问题，下面训练阶段swanlab可视化0.6B模型的输出：

Question: 1895年德国物理学教授伦琴的发现对医学影像学的发展有何具体影响？请从技术进步、学科建立和临床应用三个方面进行分析。

LLM: Description* CustomClass* Customlying Time Name * Name Full Solution Answer Solution Question Answer

• CustomestQuestionQuestionClass** ** * * * * Answer * ** Solution Time Original Solution Description Answer Custom Access Instructions Solution Answer Solution Time Full * Solution Answer Answer*** Navigation Navigation Answer* * * Solution Explanationeder** * Solution ** ** Time Solution Answer* Full Student Review* Solution * * * Navigation Question Name* - Question Instructions * Navigation * Original
• • • • • • • • Name Answer Answer Answer * * * Name Name Time* * 1 Question * ImportedestlyingClass NameedClass Question Full Name Navigation * * Time Answer Solution Description * 0 * Import* Original Question Instructions* Name * Import * Answer * Question * * - * Time * 1* Answer* *
                Name Solution* Solution Answer * * * Name: * - Customlying Question * * * * * * * * * Answer * * * * Answer** Name * * Solution * Answer Question * Answer SolutionQuestion * * * * * Time Question Answer Full Solution * * * * * * Custom Solution Access Navigation Time Question * Time * * AnswerClassest* Access* * - Answer NameQuestion * Question Name * * ** Name Solution ** Answer* Name Time 1- Access Solution TimeClass Solution * * * ** Original Answer * Solution: * * Import Solution Description Answer Name Solution * * * *

Question Solution Name Name * Student Answerested** Solution * * ** Answer* Question Instructions Name * * Time Navigation Solution 1 Answer* 0* Solution * *
Answer *** * * Question Answer * * Simple ReviewedClass * Solution
Multiple Name Question Question Question * * Navigationestly Review Time Solution 1 Question AnswerClass Answerest Solution Question * Navigation Description Answer Answer Solution Custom Review Answer Instructionsed Answer Name Navigation Navigation Answer Full Math 0 Answer Name Question Answer Solution Question Navigation Time Answer Time Original Name * * * Solution
** -* * * Custom Instructions Answer Question* Solution Navigation * * * Student Answer Instructions Fullest Question * * Time Solution ** Time Solution Custom Name* *
Question Answer* Customest Solution * Navigation Accessed Solution Answerested* Answerester Name Full QuestionClass Description *** * * * * * Solution * ** Custom * Answer Question * * Question * ** Solution Question Review Navigation Name Question Answer * * * Answer Full Name Solution CustomClass Answer Custom Name Question* * * Original Customest Answer * Custom - * Solution Navigation Answer Solution Name Review * ** QuestionQuestion Question Answer Answer *

• • • • • • Solution * * * *
• Fullest Solution: * Answer * * Answer - Math Answer Question Explanation Name Answer Navigation Custom Answer * Full MathClass Name * * * Name Custom Answer * * Question AnswerClass Answer Answer * Solution Original * Time Full Answer Answer * Time AccessClassested * * * ** Solution* * * * Question Full Review Answer Name estest Nameest * ** Full * Customly * Original Solution Question Description* Answer Question* * * Answer * * * * Simple Questionester Solution* Answer Original * * * Solution Name Custom Answer Answer ** Solution Answer
• • • Answer Answer * * ** est Solution Solution Answer Originalestest * Question Answer Answer Name Solution: Answer Answer * Question Explanation Answer * * * * Student Answer* * * * * Name: * * ** * * Solution Solution * * ** Answer Solution Question Question Solution* Question Explanation Instructions Question Solution Name * Question Answer Customestroducing Original * * * * * * - Full* * Re* * Import Solution Solution * Time Solution** * Solution estestest Answer * * * Import Solution Navigation Custom Answer* * * * Answer Question Solution** Questionestly
      estClass QuestionQuestion Full Name Solution Name Custom Solution * - Question Answer Answer Answer ** Full * Answer Question * * Solution* * * * * * * * Importlying* Question Question * * * QuestionClass * Student Full Answer Name Import Accessest Solution QuestionQuestion Answer Answer Solution * Question Answer Answer * * * * * Time Name Answer * * ** * * * Class Question * **** Solution Solution* *
• Solution *Question Answer * Time Time Question Question Description
• Answer Original Answerest Answer Custom Solution Solution Explanation Answer Solution Solution Answer Time * Answer Answer Solution * Relying Questionestest Answer * * Question Fullest Review Answer Time* Custom Name * * * Question Instructions Navigation * Student Student Navigation Solution 0 1 * * ** * Solution: * Question**
• Answer * * QuestionClass Solution 이 Answer Custom * * Navigation Solution- Student Original 0
  Solution: Navigation * Class Solution*Class -Question Navigation Question Solution Review 11 * Originalester ** Solution Name * *
• ** - Answer * * Question Question Question Explanation Answer Customlyestroducing * * AnswerQuestion Description Answer Question Question ** 3 Review Questionest Answer
  Question Student Full Student Time Question * ** * Navigation Full Name NavigationClass Explanation- Name Question Answerestlyingly

• Solution Navigation Solution Answer * Import * Student Solution Navigation Navigation Name * * * * Solution Answer Answer Solution * * Original Navigation * * Name * * Solution* ** ** Name * SolutionQuestion: Access * * Import Student Answer ** ** 3Question Question Answer Answer Solution * Solution Solution * Solution * * Custom Reviewestestroducing Answer Name * * Questionester Solution * * * * * Answer * Question ReviewQuestion * * * - Answer Solution Solution Solution Name * Custom 1 * Full Fullly * * * * Student Student**Class Explanation Instructions* Solution* Question * *

• • • Answer Answer * Import Solution * Solution Answer - * ** Class ** Text * * Answer Answer Answer Time Answer Answer
      QuestionClass Question Question Full 1 Answer Answer Answer Description * Import Question Answer * Time Explanation Solution Explanation Name Explanation Full * ** * Solution * * Question Navigationest * * QuestionQuestion Nameest Answer * Name Answer Question Answer
      Question * Question Solution Solution *
      Classroducing Answer Instructions Instructions ** ** Solution Time Question Name Full Answer
      Solution Explanation** Full* Answer Answer Name Solutionestestest Answer Full AnswerQuestion Answer Question Name * ** Solution Time Description Name Answer Custom Question * ** Accessestles Solution Navigation Navigation* Solution Navigation * Time Simplely Answer* Class ** Navigation Custom Review Navigation Answer Description Name Question Instructions Instructions Review Answer * * * * * Answer Name Student Original Navigation Answer* Solution * * * Time Reviewestlying * Name: *- 10! * Original Answer Answer Custom Name Solution Navigation * Answer Answer Description Solution * * Solution: Question Explanation- Student Name Solution * Question Import * * * * Answer * * Full Question Question Question Instructions
      est Solution * * * * Original Answer Navigation Navigation * * Student Custom Solution ExplanationClass Solution *
• • Solution Explanation Solution Navigation * Question Description Custom Review * Question ** - Answer Navigation Question Answer * * AnswerClass Explanation: Name Custom Description Customest * QuestionClass Answer * Answer Answer Explanation * * * Student * * Question Full Solution Question Time*- * * Original * * * * Classest Solution Explanation Navigation Instructions * * * Import Solution ReviewClass * Question * - Answer * * * * * - * *
    Question* * ImportClassroducing Original Answer Question Solution Answer Time Original * Solution * * Name Solution * * Solution Answer * * Navigation ** Solution * * Answer Solution Navigation NavigationClass Instructions * * ** - Access * * * * * Import Solution Name Full Answer Original Solution* Custom Reviewest Question Name Student Navigation** * ** Access Name * Name Question Answer* * Question Questionestlying Navigation Instructions * Questionestestest * * * Solution Timeest * Answer Name Custom Custom * Custom Solution Solution Answer Name Explanationest Solution: Full Customestroducing Time Instructions Question Question* Solution * * * Question Name Import* Answer Navigationlyestroducing Question Question Review Answerest Name *

[ppppps]

啊我也遇到了这个问题，下面训练阶段swanlab可视化0.6B模型的输出：

Question: 1895年德国物理学教授伦琴的发现对医学影像学的发展有何具体影响？请从技术进步、学科建立和临床应用三个方面进行分析。

LLM: Description* CustomClass* Customlying Time Name * Name Full Solution Answer Solution Question Answer

• CustomestQuestion_QuestionClass** ** * * * * Answer * ** Solution Time Original Solution Description Answer_ Custom Access Instructions Solution Answer Solution Time Full * Solution Answer Answer*** Navigation Navigation Answer* * * Solution Explanationeder** * Solution ** ** Time Solution Answer* Full Student Review* Solution * * * Navigation Question Name* - Question Instructions * Navigation * Original

• • • • • • • • Name Answer Answer Answer * * * Name Name Time* * 1 Question * ImportedestlyingClass NameedClass Question Full Name Navigation * * Time Answer Solution Description * 0 * Import* Original Question Instructions* Name * Import * Answer * Question * * - * Time * 1* Answer* *
                Name Solution* Solution Answer * * * Name: * - Customlying Question * * * * * * * * * Answer * * * * Answer** Name * * Solution * Answer Question * Answer Solution_Question * * * * * Time Question Answer Full Solution * * * * * * Custom Solution_ Access Navigation Time Question * Time * * AnswerClassest* Access* * _- Answer Name_Question * Question Name * * ** Name Solution ** Answer* Name Time 1- Access Solution TimeClass Solution * * * ** Original Answer * Solution: * * Import Solution Description Answer Name Solution * * * *

Question Solution Name Name * Student Answerested** Solution * * ** Answer* Question Instructions Name * * Time Navigation Solution 1 Answer* 0* Solution * * Answer *** * * Question Answer * * Simple ReviewedClass * Solution Multiple Name Question Question Question * * Navigationestly Review Time Solution 1 Question AnswerClass Answerest Solution Question * Navigation Description Answer Answer Solution Custom Review Answer Instructionsed Answer Name Navigation Navigation Answer Full Math 0 Answer Name Question Answer Solution Question Navigation Time Answer Time Original Name * * * Solution ** -* * * Custom Instructions Answer Question* Solution Navigation * * * Student Answer Instructions Fullest Question * * Time Solution ** Time Solution Custom Name* * Question Answer* Customest Solution * Navigation Accessed Solution Answerested* Answerester Name Full QuestionClass Description *** * * * * * Solution * ** Custom * Answer Question * * Question * ** Solution Question Review Navigation Name Question Answer * * * Answer Full Name Solution CustomClass Answer Custom Name Question* * * Original Customest Answer * Custom - * Solution Navigation Answer Solution Name Review * ** QuestionQuestion Question Answer Answer *

• • • • • • Solution * * * *

• Fullest Solution: * Answer * * Answer - Math Answer Question Explanation Name Answer Navigation Custom Answer * Full MathClass Name * * * Name Custom Answer * * Question AnswerClass Answer Answer * Solution Original * Time Full Answer Answer * Time AccessClassested * * * ** Solution* * * * Question Full Review Answer Name estest Nameest * ** Full * Customly * Original Solution Question Description* Answer Question* * * Answer * * * * Simple Questionester Solution* Answer Original * * * Solution Name Custom Answer Answer ** Solution Answer

• • • Answer Answer * * ** est Solution Solution Answer Originalestest * Question Answer Answer Name Solution: Answer Answer * Question Explanation Answer * * * * Student Answer* * * * * Name: * * ** * * Solution Solution * * ** Answer Solution Question Question Solution* Question Explanation Instructions Question Solution Name * Question Answer Customestroducing Original * * * * * * - Full* * Re* * Import Solution Solution * Time Solution** * Solution estestest Answer * * * Import Solution Navigation Custom Answer* * * * Answer Question Solution** Questionestly
      estClass QuestionQuestion Full Name Solution Name Custom Solution * - Question Answer Answer Answer ** Full * Answer Question * * Solution* * * * * * * * Importlying* Question Question * * * QuestionClass * Student Full Answer Name Import Accessest Solution QuestionQuestion Answer Answer Solution * Question Answer Answer * * * * * Time Name Answer * * ** * * * Class Question * **** Solution Solution* *

• Solution *Question Answer * Time Time Question Question Description

• Answer Original Answerest Answer Custom Solution Solution Explanation Answer Solution Solution Answer Time * Answer Answer Solution * Relying Questionestest Answer * * Question Fullest Review Answer Time* Custom Name * * * Question Instructions Navigation * Student Student Navigation Solution 0 1 * * ** * Solution: * Question**

• Answer * * QuestionClass Solution 이 Answer Custom * * Navigation Solution- Student Original 0
  Solution: Navigation * Class Solution*Class -Question Navigation Question Solution Review 11 * Originalester ** Solution Name * *

• ** - Answer * * Question Question Question Explanation Answer Customlyestroducing * * Answer_Question Description Answer Question Question ** 3 Review Questionest Answer
  Question Student Full Student Time Question * ** * Navigation Full Name Navigation_Class Explanation- Name Question Answerestlyingly

• Solution Navigation Solution Answer * Import * Student Solution Navigation Navigation Name * * * * Solution Answer Answer Solution * * Original Navigation * * Name * * Solution* ** ** Name * Solution_Question: Access * * Import Student Answer ** ** 3Question Question Answer Answer Solution * Solution Solution * Solution * * Custom Reviewestestroducing Answer Name * * Questionester Solution * * * * * Answer * Question ReviewQuestion * * * - Answer Solution Solution Solution Name * Custom 1_ * Full Fullly * * * * Student Student**Class Explanation Instructions* Solution* Question * *

• • • Answer Answer * Import Solution * Solution Answer - * ** Class ** Text * * Answer Answer Answer Time Answer Answer
      QuestionClass Question Question Full 1 Answer Answer Answer Description * Import Question Answer * Time Explanation Solution Explanation Name Explanation Full * ** * Solution * * Question Navigationest * * QuestionQuestion Nameest Answer * Name Answer Question Answer
      Question * Question Solution Solution *
      Classroducing Answer Instructions Instructions ** ** Solution Time Question Name Full Answer
      Solution Explanation** Full* Answer Answer Name Solution_estestest Answer_ Full Answer_Question Answer Question_ Name * ** Solution Time Description Name Answer Custom Question * ** Accessestles Solution Navigation Navigation* Solution Navigation * Time Simplely Answer* Class ** Navigation Custom Review Navigation Answer Description Name Question Instructions Instructions Review Answer * * * * * Answer Name Student Original Navigation Answer* Solution * * * Time Reviewestlying * Name: *- 10! * Original Answer Answer Custom Name Solution Navigation * Answer Answer Description Solution * * Solution: Question Explanation- Student Name Solution * Question Import * * * * Answer * * Full Question Question Question Instructions
      est Solution * * * * Original Answer Navigation Navigation * * Student Custom Solution ExplanationClass Solution *
• • Solution Explanation Solution Navigation * Question Description Custom Review * Question ** - Answer Navigation Question Answer * * AnswerClass Explanation: Name Custom Description Customest * QuestionClass Answer * Answer Answer Explanation * * * Student * * Question Full Solution Question Time*- * * Original * * * * Classest Solution Explanation Navigation Instructions * * * Import Solution ReviewClass * Question * - Answer * * * * * - * *
    Question* * Import_Classroducing Original Answer Question Solution Answer Time Original * Solution * * Name Solution * * Solution Answer * * Navigation ** Solution * * Answer Solution Navigation Navigation_Class Instructions * * ** - Access * * * * * Import Solution Name Full Answer Original Solution* Custom Reviewest Question Name Student Navigation** * ** Access Name * Name Question Answer* * Question Question_estlying Navigation Instructions * Questionestestest * * * Solution Timeest * Answer Name Custom Custom * Custom Solution_ Solution Answer Name Explanationest Solution: Full Customestroducing Time Instructions Question Question* Solution * * * Question Name Import* Answer Navigationlyestroducing Question Question Review Answerest Name *

问题好像解决了，在train.py中修改了predict函数：

def predict(messages, model, tokenizer):
device = "cuda" if torch.cuda.is_available() else "cpu"
# 切换到评估模式
model.eval()
text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)
# 使用torch.no_grad()减少内存使用
with torch.no_grad():
generated_ids = model.generate(
model_inputs.input_ids,
max_new_tokens=2048, # 增加生成长度
temperature=0.7, # 添加温度参数控制随机性
top_p=0.95, # 添加top-p采样
do_sample=True, # 启用采样
num_return_sequences=1
)
# 切回训练模式
model.train()
generated_ids = [
output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]
response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
return response

[skyler9901]

请教一个问题，为什么text的tokenizer要在评估模式下操作，generate_id要在训练模式生成？

[CarrotSwordsman]

乱码是由于在推理的时候仍在训练模式，没有切换到评估模式导致。在源代码的predict中添加两行model.eval()和model.train()即可

def predict(messages, model, tokenizer):
    device = "cuda"

    model.eval()  # 关键：切换到评估模式
    
    text = tokenizer.apply_chat_template(
        messages,
        tokenize=False,
        add_generation_prompt=True
    )
    model_inputs = tokenizer([text], return_tensors="pt").to(device)

    generated_ids = model.generate(
        model_inputs.input_ids,
        max_new_tokens=MAX_LENGTH,
    )
    
    model.train()  # 切回训练模式
    
    generated_ids = [
        output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
    ]

    response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

    return response```