这可能是全网最保姆的DeepSeek微调教学

from unsloth import FastLanguageModel

max _ seq _ length  = 2048 dtype  = None load _ in _ 4bit  = False

FastLanguageModel . for _ inference  ( model  )

question  =  " 请问如何证明根号 2 是无理数？ "

inputs  = tokenizer  (  [  question  ] , return _ tensors=  "pt"  ).to  (  "cuda"  )

outputs  = model.generate  ( input _ ids  = inputs.input _ ids ,max _ new _ tokens  = 1200 , use _ cache  = True ,  )

response  = tokenizer.batch _ decode ( outputs )

response

print ( response [ 0 ])

prompt _ style = """Below is an instruction that describes a task , paired with an input that provides further context.Write a response that appropriately completes the request.Before answering , think carefully about the question and create a step- by  -step chain of thoughts to ensure a logical and accurate response.
 ### Instruction : You are a medical expert with advanced knowledge in clinical reasoning , diagnostics , and treatment planning.Please answer the following medical question.
 ### Question : { }
 ### Response :  <think> { } """ 

question _ 1 = "A 61-year-old woman with a long history of involuntary urine loss during activities like coughing or sneezing but no leakage at night undergoes a gynecological exam and Q-tip test.Based on these findings , what would cystometry most likely reveal about her residual volume and detrusor contractions?  " 

question _ 2 = "Given a patient who experiences sudden-onset chest pain radiating to the neck and left arm , with a past medical history of hypercholesterolemia and coronary artery disease , elevated troponin I levels , and tachycardia , what is the most likely coronary artery involved based on this presentation?  " 

inputs1 = tokenizer ([ prompt _ style.format ( question _ 1 , ""  )  ] , return _ tensors=  "pt"  ).to  (  "cuda"  ) 

outputs1 = model.generate  ( input _ ids=inputs1.input _ ids ,max _ new _ tokens=1200 , use _ cache=True ,  ) 
response1 = tokenizer.batch _ decode ( outputs1 )

print  ( response1  [  0  ] . split  (  "### Response : "  )  [  1  ]  )

 <think> Okay , so I  'm trying to figure out this medical question.Let' s break it down step by step.The patient is a 61  -year-old woman with a history of involuntary urine loss during activities like coughing or sneezing , but she doesn  't leak at night.She' s undergoing a gynecological exam and a Q-tip test.The question is about what cystometry would show regarding her residual volume and detrusor contractions.
First , I know that involuntary urine loss during activities like sneezing or coughing is often related to stress urinary incontinence.Stress incontinence usually happens when the urethral muscles aren  't strong enough to prevent the bladder from leaking when there' s increased pressure , like from coughing .
Now , the Q-tip test.From what I remember , the Q-tip is a urethral catheter used to measure urethral pressure profiles.It  's often used to assess urethral function.A positive Q-tip test result , where the urethral pressure is lower than the intravesicular pressure during the Valsalva maneuver , is associated with intrinsic sphincteric deficiency , which is a type of stress incontinence .
Since the patient has a history of involuntary leakage during activities but not at night , it' s more likely stress incontinence rather than something like urgency incontinence , which would also involve nighttime leakage.So ,if the Q-tip test is positive , it points towards intrinsic sphincteric deficiency .
Now , moving to cystometry.Cystometry is a test that measures how the bladder responds to filling and the contractions of the detrusor muscle.It can show if there  's an overactive bladder ( OAB ), which causes urgency and frequency , or if there' s underactive detrusor , which can lead to retention .
In this case , the patient  's main issue is stress incontinence , which is more about the inability to hold urine under increased pressure.Cystometry would look at the detrusor contractions.If the detrusor is underactive , it won' t contract strongly enough to empty the bladder , leading to residual volume.But if it  's overactive , it might contract too strongly , causing urgency .
Given that the patient has a history of stress incontinence and a positive Q-tip test suggesting intrinsic sphincteric deficiency , I think the cystometry would show that the detrusor contractions are normal.The issue isn' t with the detrusor muscle  's ability to contract but rather the inability to seal the urethra to hold the pressure.So , the residual volume might be normal unless there' s significant retention , but the key finding would be the detrusor contractions being normal ,not overactive.
Wait , but could there be residual volume? If the patient has some urine left in the bladder after urination , that  's residual volume.But without symptoms of retention , like a full bladder or difficulty starting urination , it' s less likely.The main issue is the incontinence during activities , so the detrusor contractions are normal ,and the residual volume is within normal limits unless there  's another condition.
So , putting it all together , the cystometry would likely show normal detrusor contractions and normal residual volume.The issue is more on the sphincteric side , not the detrusor . </think> 
Basedontheanalysisofthepatient' shistory and theQ-tiptestresults , thecystometrywouldmostlikelyrevealnormaldetrusorcontractions and normalresidualvolume.Theprimaryissueappearstobestressurinaryincontinenceduetointrinsicsphinctericdeficiency , asindicatedbythepositiveQ-tiptest.Thisconditiontypicallyaffectstheurethralsphincter  'sabilitytopreventleakageunderincreasedpressure , ratherthanthedetrusormuscle' scontractility.Therefore , thedetrusorcontractionsare not overactive , and theresidualvolumeiswithinnormallimits.< ｜end ▁ of ▁ sentence ｜ > 

inputs2 = tokenizer ([ prompt _ style.format ( question _ 2 , ""  )  ] , return _ tensors=  "pt"  ).to  (  "cuda"  ) 

outputs2 = model.generate  ( input _ ids=inputs2.input _ ids ,max _ new _ tokens=1200 , use _ cache=True ,  ) 
response2 = tokenizer.batch _ decode ( outputs2 )

print  ( response2  [  0  ] . split  (  "### Response : "  )  [  1  ]  )

 <think> Okay , so I have a patient presenting with sudden-onset chest pain that's radiating to the neck and left arm.That makes me think of a heart attack because those symptoms are classic — angina or myocardial infarction.The left arm pain , neck pain , and sometimes jaw or back pain can all be related to coronary artery issues .
Looking at the past medical history , the patient has hypercholesterolemia , which is high cholesterol , and coronary artery disease .Those are both risk factors for atherosclerosis , which can lead to blockages in the coronary arteries.The elevated troponin I levels are a big clue because troponin is a cardiac enzyme released when the heart muscle is damaged , which is a sign of a heart attack.Also , the patient is experiencing tachycardia , which means their heart is beating faster than usual.In a heart attack , the heart might beat faster as it tries to pump blood to compensate for the blocked artery.
Now , considering the coronary arteries , the left main coronary artery supplies blood to the entire left side of the heart , including the left ventricle , which is a large muscle that's crucial for pumping blood.If there's a blockage here , it can lead to a more severe heart attack because the left ventricle is so vital.The right coronary artery supplies the right ventricle and the inferior wall of the left ventricle.Blockages here are possible too , but the left main is more commonly associated with the symptoms described , especially when troponin is elevated.
So putting it all together , the most likely coronary artery involved is the left main coronary artery.The combination of the patient's history , the elevated troponin , and the typical chest pain radiation points to this artery being the culprit . </think> 
The most likely coronary artery involved in this presentation is the ** left main coronary artery  ( LMCA  )  **.
 ** Explanation : **  - ** Symptoms : ** The patient's sudden chest pain radiating to the neck and left arm , along with elevated troponin levels , suggests an acute coronary syndrome , likely a myocardial infarction ( heart attack ). - ** Past Medical History : ** History of hypercholesterolemia and coronary artery disease are risk factors for atherosclerosis , which can lead to blockages in the coronary arteries. - ** Tachycardia : ** Increased heart rate may occur as the heart compensates for reduced blood flow to the heart muscle. - ** Coronary Artery Consideration : ** The left main coronary artery supplies the left ventricle , a large muscle that is crucial for cardiac function.Blockages in the LMCA can lead to more severe and life-threatening heart attacks compared to blockages in the right coronary artery , which typically supply less critical areas .
Thus , thecombinationofsymptoms , elevatedtroponin , andthepatient'shistorystronglypointstothe** left maincoronaryartery** as themostlikelyculprit.< ｜ end ▁ of ▁ sentence ｜ > 

import subprocess import os 
result = subprocess.run ( 'bash - c  "source /etc/network _ turbo && env | grep proxy"  ' , shell= True , capture _ output= True , text= True  ) output = result.stdout for line in output.splitlines  (  ) :if  '=' in line : var , value = line .split  (  '=' , 1  ) os.environ  [ var  ]  = value 

！pip install datasets

import os from datasets import load _ dataset

EOS _ TOKEN = tokenizer.eos _ token tokenizer.eos _ token

 '< ｜ end ▁ of ▁ sentence ｜ >'

def formatting _ prompts _ func  (  examples  ) : inputs = examples  [  "Question"  ] cots = examples  [  "Complex _ CoT"  ] outputs = examples  [  "Response"  ] texts =  [  ] for input , cot , output in zip ( inputs , cots , outputs ):text = train _ prompt _ style.format ( input , cot , output ) + EOS _ TOKEN texts.append ( text )return {  "text" : texts , }

dataset  =load _ dataset  (  "FreedomIntelligence/medical-o1-reasoning-SFT" ,  "en" , split=  "train [ 0 : 500 ] " , trust _ remote _ code= True  )

Using the latest cached version of the dataset since FreedomIntelligence/medical-o1-reasoning-SFT couldn't be found on the Hugging Face Hub Found thelatestcacheddatasetconfiguration'en'at/root/.cache/huggingface/datasets/FreedomIntelligence ___ medical-o1-reasoning-sft/en/0.0.0/4c9573e7de1e8660b88158db2efa7c7204bbd269 ( lastmodified on WedFeb501 : 06 : 322025  ) .

dataset  [  0  ]

{  'Question' :  'A 61-year-old woman with a long history of involuntary urine loss during activities like coughing or sneezing but no leakage at night undergoes a gynecological exam and Q-tip test.Based on these findings , what would cystometry most likely reveal about her residual volume and detrusor contractions?' ,  'Complex _ CoT' :  "Okay , let's think about this step by step.There's a 61-year-old woman here who's been dealing with involuntary urine leakages whenever she's doing something that ups her abdominal pressure like coughing or sneezing.This sounds a lot like stress urinary incontinence to me.Now , it's interesting that she doesn't have any issues at night; she isn't experiencing leakage while sleeping.This likely means her bladder's ability to hold urine is fine when she isn't under physical stress.Hmm , that's a clue that we're dealing with something related to pressure rather than a bladder muscle problem.\n\nThe fact that she underwent a Q-tip test is intriguing too.This test is usually done to assess urethral mobility.In stress incontinence , a Q-tip might move significantly , showing urethral hypermobility.This kind of movement often means there's a weakness in the support structures that should help keep the urethra closed during increases in abdominal pressure.So , that's aligning well with stress incontinence.\n\nNow , let's think about what would happen during cystometry.Since stress incontinence isn't usually about sudden bladder contractions , I wouldn't expect to see involuntary detrusor contractions during this test.Her bladder isn't spasming or anything; it's more about the support structure failing under stress.Plus , she likely empties her bladder completely because stress incontinence doesn't typically involve incomplete emptying.So , her residual volume should be pretty normal.\n\nAll in all , it seems like if they do a cystometry on her , it will likely show a normal residual volume and no involuntary contractions.Yup , I think that makes sense given her symptoms and the typical presentations of stress urinary incontinence." ,  'Response' :  'Cystometryinthiscaseofstressurinaryincontinencewouldmostlikelyrevealanormalpost-voidresidualvolume , asstressincontinencetypicallydoesnotinvolveissueswithbladderemptying.Additionally , sincestressurinaryincontinenceisprimarilyrelatedtophysicalexertionandnotanoveractivebladder , youwouldnotexpecttoseeanyinvoluntarydetrusorcontractionsduringthetest.'  }

dataset  =dataset.map ( formatting _ prompts _ func , batched= True ,  )

dataset  [  "text"  ]  [ 0  ]

model = FastLanguageModel.get _ peft _ model  ( model , r= 16 , target _ modules=  [  "q _ proj" ,  "k _ proj" ,  "v _ proj" ,  "o _ proj" ,  "gate _ proj" ,  "up _ proj" ,  "down _ proj" ,  ] , lora _ alpha= 16 , lora _ dropout= 0 , bias=  "none" , use _ gradient _ checkpointing=  "unsloth" , # True or  "unsloth" for very long context random _ state= 3407 , use _ rslora=False , loftq _ config=None ,  )

Unsloth 2025 . 1 . 8 patched 32 layers with 32 QKV layers , 32 O layers and 32 MLP layers . 

from trl import SFTTrainer from transformers import TrainingArguments from unsloth import is _ bfloat16 _ supported

trainer  = SFTTrainer  ( model  = model , tokenizer  = tokenizer , train _ dataset  = dataset , dataset _ text _ field  =  "text" , max _ seq _ length  = max _ seq _ length , dataset _ num _ proc  = 2 , args  = TrainingArguments  ( per _ device _ train _ batch _ size  = 2 , gradient _ accumulation _ steps  = 4 ,  # Use num _ train _ epochs = 1 , warmup _ ratio for full training runs! warmup _ steps  = 5 , max _ steps  = 60 , learning _ rate  = 2e-4 , fp16  = not is _ bfloat16 _ supported  (  ) , bf16  = is _ bfloat16 _ supported  (  ) , logging _ steps  = 10 , optim  =  "adamw _ 8bit" , weight _ decay  = 0.01 ,lr _ scheduler _ type  =  "linear" , seed  = 3407 , output _ dir  =  "outputs" ,  ) ,  )

trainer _ stats  = trainer.train  (  )

 /root/autodl-tmp/models/wandb/run-20250205 _ 004957-k0dz6rg7

trainer _ stats

FastLanguageModel . for _ inference  ( model  )

inputs = tokenizer ([ prompt _ style.format ( question _ 1 , ""  )  ] , return _ tensors=  "pt"  ).to  (  "cuda"  ) 
outputs = model.generate  ( input _ ids=inputs.input _ ids ,attention _ mask=inputs.attention _ mask ,max _ new _ tokens=1200 , use _ cache=True ,  ) response = tokenizer.batch _ decode ( outputs )

print  ( response  [  0  ] . split  (  "### Response : "  )  [  1  ]  )

new _ model _ local =  "DeepSeek-R1-Medical-COT-Tiny" model.save _ pretrained ( new _ model _ local )tokenizer.save _ pretrained ( new _ model _ local )
model.save _ pretrained _ merged ( new _ model _ local , tokenizer , save _ method =  "merged _ 16bit" ,  ) 

train _ prompt _ style = """Below is an instruction that describes a task , paired with an input that provides further context.Write a response that appropriately completes the request.Before answering , think carefully about the question and create a step- by  -step chain of thoughts to ensure a logical and accurate response.
 ### Instruction : You are a medical expert with advanced knowledge in clinical reasoning , diagnostics , and treatment planning.Please answer the following medical question.
 ### Question : { }
 ### Response :  <think> { } </think> { } """ 

EOS _ TOKEN = tokenizer.eos _ token  # Must add EOS _ TOKEN 

def formatting _ prompts _ func  (  examples  ) : inputs = examples  [  "Question"  ] cots = examples  [  "Complex _ CoT"  ] outputs = examples  [  "Response"  ] texts =  [  ] for input , cot , output in zip ( inputs , cots , outputs ):text = train _ prompt _ style.format ( input , cot , output ) + EOS _ TOKEN texts.append ( text )return {  "text" : texts , }

dataset = load _ dataset ( "FreedomIntelligence/medical-o1-reasoning-SFT" , "en" , split = "train" , trust _remote _ code=True  ) dataset = dataset.map ( formatting _ prompts _func , batched = True , ) dataset  [  "text"  ]  [ 0  ]

model = FastLanguageModel.get _ peft _ model  ( model , r= 16 , target _ modules=  [  "q _ proj" ,  "k _ proj" ,  "v _ proj" ,  "o _ proj" ,  "gate _ proj" ,  "up _ proj" ,  "down _ proj" ,  ] , lora _ alpha= 16 , lora _ dropout= 0 , bias=  "none" , use _ gradient _ checkpointing=  "unsloth" , # True or  "unsloth" for very long context random _ state= 3407 , use _ rslora=False , loftq _ config=None ,  )

from trl import SFTTrainer from transformers import TrainingArguments from unsloth import is _ bfloat16 _ supported 
trainer  = SFTTrainer  ( model  = model , tokenizer  = tokenizer , train _ dataset  = dataset , dataset _ text _ field  =  "text" , max _ seq _ length  = max _ seq _ length , dataset _ num _ proc  = 2 , args  = TrainingArguments  ( per _ device _ train _ batch _ size  = 2 , gradient _ accumulation _ steps  = 4 , num _ train _ epochs  = 3 , warmup _ steps  = 5 ,  # max _ steps=60 , learning _ rate  = 2e-4 , fp16  = not is _ bfloat16 _ supported  (  ) , bf16  = is _ bfloat16 _ supported  (  ) , logging _ steps  = 10 , optim  =  "adamw _ 8bit" , weight _ decay  = 0.01 ,lr _ scheduler _ type  =  "linear" , seed  = 3407 , output _ dir  =  "outputs" ,  ) ,  ) 

trainer _ stats  = trainer.train  (  )

 <think> Alright , let  's think this through step by step.We' ve got a 61  -year-old woman who  's been dealing with involuntary urine loss whenever she does something like coughing or sneezing.That sounds like stress urinary incontinence , which usually means there' s some kind of leakage when there  's increased pressure in the abdomen.Now , the fact that she doesn' t have any leakage at night is interesting.That  's a big hint because it suggests that the problem isn' t about bladder control or retention , since there  's no issue when lying down.
Now , let' s consider the Q-tip test.This test is a clever way to check for urethral mobility.If the Q-tip moves a lot , it means the urethra is mobile , which is a typical sign of stress incontinence.So , we  're probably looking at a condition where the urethra isn' t staying closed under pressure.
Now , if we were to do a cystometry on her , we  'd be checking a few things.First , there' s the residual volume.Since she  's not leaking urine at night , her bladder should be empty , so we' d expect the residual volume to be normal.Then , there  's the detrusor contractions.Stress incontinence doesn' t usually involve the detrusor muscle contracting abnormally.So , we  'd expect normal or maybe slightly increased contractions , but nothing out of the ordinary .
So , putting it all together , if we do the cystometry , we' d likely see a normal residual volume and normal detrusor contractions.That fits with everything we know about her symptoms and the Q-tip test results.Everything lines up with stress urinary incontinence , confirming that our conclusion makes sense . </think> Based on the information provided , the 61  -year-old woman is likely experiencing stress urinary incontinence , as suggested by the involuntary urine loss during activities like coughing or sneezing , coupled with the fact that she doesn  't experience leakage at night.During a cystometry exam , we would expect the following findings :
1.**Residual Volume : ** This would be normal.Since she does not experience leakage at night , her bladder is likely empty at bedtime , indicating a normal post-void residual volume .
2.**Detrusor Contractions : ** These would also be normal or possibly slightly increased.Stress urinary incontinence typically does not involve abnormal detrusor contractions , as the issue is more related to the closure mechanism of the urethra under increased abdominal pressure rather than bladder control or contractions .
Thus , cystometry would most likely reveal a normal residual volume and normal detrusor contractions , consistent with stress urinary incontinence.< ｜ end ▁ of ▁ sentence ｜ > 

 <think> Okay , let  's think about this.The patient is having sudden chest pain that's shooting to their neck and left arm.That's pretty classic for something like a myocardial infarction , right  ? I mean , the pain spreading like that usually points to a big problem in the heart's supply.
Now , considering the heart's coronary arteries ,I know the left anterior descending artery , or LAD , is a big player.It's like the main highway for blood to reach the front wall of the heart.If there's a blockage there , it can definitely cause pain that radiates to the neck and arm .
Then , there's the right coronary artery , or RCA , which supplies the right side of the heart and can affect the inferior wall of the heart.But wait , the pain pattern here seems to be more on the left side , so maybe the RCA is less likely.
The patient has hypercholesterolemia and coronary artery disease.These conditions put them at risk for atherosclerosis , which can lead to blockages in the coronary arteries.The LAD is commonly involved in such scenarios , especially when the pain spreads to the neck and arm .
Also , the elevated troponin I levels and tachycardia are strong signals that something serious is happening in the heart.These are usually seen in myocardial infarctions.Given the pain pattern and the patient's risk factors , the LAD seems like the most likely culprit here.
So , when I put all this together , it really seems like the left anterior descending artery is the most likely artery involved in this situation.It just fits with the classic presentation of anterior myocardial infarction.Yeah , I'm pretty confident about that. </think> Based on the presentation of sudden-onset chest pain radiating to the neck and left arm , along with the patient's history of hypercholesterolemia and coronary artery disease , the most likely coronary artery involved is the left anterior descending  ( LAD  )  artery.This artery supplies the front wall of the heart , and a blockage here can cause the classic symptoms described .The elevated troponin I levels and tachycardia further support the likelihood of a myocardial infarction , with the LAD being a common site for such events.< ｜ end ▁ of ▁ sentence ｜ > 

new _ model _ local =  "DeepSeek-R1-Medical-COT" model.save _ pretrained ( new _ model _ local )tokenizer.save _ pretrained ( new _ model _ local )
model.save _ pretrained _ merged ( new _ model _ local , tokenizer , save _ method=  "merged _ 16bit" ,  )