
ããã³ããã®å ãž â èªåã®ã¢ãã«ããã¡ã€ã³ãã¥ãŒãã³ã°ããçç±ã𿹿³
ãã¡ã€ã³ãã¥ãŒãã³ã°ã¯AI婿޻çšãé²ãã§ãäŒæ¥ã»ã©ããŸã䜿ã£ãŠãå°è±¡
LLMã¯ãæ±çšçãªç¥èœããã¯ãããã«ããã¯ã§ã¯ãªã段éã«å°éããŸããããšã³ã¿ãŒãã©ã€ãº AI ã·ã¹ãã ã«ãããæ¬åœã®èª²é¡ã¯ãè¡åã®ã¢ã©ã€ã³ã¡ã³ãïŒbehavioral alignmentïŒã§ããã€ãŸããå€§èŠæš¡ç°å¢ã«ãããŠãäžè²«æ§ããããä¿¡é Œã§ããããªã·ãŒã«æºæ ããåºåãçæããã¢ãã«ã確å®ã«ããããšã§ããããã³ãããšã³ãžãã¢ãªã³ã°ãš Retrieval-Augmented GenerationïŒRAGïŒã¯åŒ·åã§ãããã¢ãã«ã®è¡åãå€ãããã®ã§ã¯ãããŸããããã¡ã€ã³ãã¥ãŒãã³ã°ã¯ãäºååŠç¿æžã¿ AI ã¢ãã«ãç¹å®ã®ã¿ã¹ã¯ãããŒã¿ã»ããã«å¯Ÿãã远å åŠç¿ã§ã«ã¹ã¿ãã€ãºããæ§èœåäžãæ°ããã¹ãã«è¿œå ã粟床åäžãè¡ãããšã§ãããã解決ããŸãã
ãã®æçš¿ã§ã¯ãMicrosoft Foundry ã®ãã¡ã€ã³ãã¥ãŒãã³ã°ãšã¯äœãããã€äœ¿ãã¹ããããµããŒãããŠãããã¡ã€ã³ãã¥ãŒãã³ã°ææ³ãããã³ Microsoft Foundry äžã§ãã¡ã€ã³ãã¥ãŒãã³ã°ãå®è¡ããããã®ã³ãŒãäŸã玹ä»ããŸãã
Microsoft Foundry ã®ãã¡ã€ã³ãã¥ãŒãã³ã°ãšã¯
Microsoft Foundry ã®ãã¡ã€ã³ãã¥ãŒãã³ã°ã䜿çšãããšãã¿ã¹ã¯åºæã®ããŒã¿ã»ããã§äºååŠç¿æžã¿ã®åºç€ã¢ãã«ïŒOpenAI ããã³ãªãŒãã³ã¢ãã«ïŒãã«ã¹ã¿ãã€ãºã§ããŸããããã«ããããŠãŒã¹ã±ãŒã¹ã«å¯ŸããŠäºæž¬å¯èœã«æ¯ãèãå°çšã¢ãã«ãäœæãã€ã€ãAzure ã®ãšã³ã¿ãŒãã©ã€ãºã°ã¬ãŒãã®ã»ãã¥ãªãã£ãã¬ããã³ã¹ãå¯èŠ³æž¬æ§ãç¶æã§ããŸãã
ãã¡ã€ã³ãã¥ãŒãã³ã°ã®äž»ãªå©ç¹ãšä»£è¡šçãªãŠãŒã¹ã±ãŒã¹

ãã¡ã€ã³ç¹åïŒDomain SpecializationïŒ: å»çãéèãæ³åŸãªã©ã®å°éé åã«èšèªã¢ãã«ãé©å¿ãããæè¡çšèªãçè§£ããããæ£ç¢ºã§ãã¡ã€ã³åºæã®åçãæäŸã§ããããã«ããã
ã¿ã¹ã¯æ§èœïŒTask PerformanceïŒ: ææ åæãã³ãŒãçæã翻蚳ãèŠçŽãªã©ã®ã¿ã¹ã¯åãã«ã¢ãã«ãæé©åããæ±çšã¢ãã«ããé«ãæ§èœãå®çŸããã
ã¹ã¿ã€ã«ãšããŒã³ïŒStyle and ToneïŒ: ãã©ãŒãã«ãªããžãã¹ããã©ã³ããã€ã¹ããã¯ãã«ã«ã©ã€ãã£ã³ã°ãªã©ã奜ã¿ã®ã³ãã¥ãã±ãŒã·ã§ã³ã¹ã¿ã€ã«ã«åãããŠã¢ãã«ããã¡ã€ã³ãã¥ãŒãã³ã°ããã
æç€ºè¿œåŸïŒInstruction FollowingïŒ: æžåŒã«ãŒã«ãè€æ°ã¹ãããã®æç€ºãæ§é ååºåãžã®è¿œåŸèœåãé«ããïŒãã«ããšãŒãžã§ã³ãã»ã¯ãŒã¯ãããŒã«ãããé©åãªãšãŒãžã§ã³ãéžæãå«ãïŒã
ã³ã³ãã©ã€ã¢ã³ã¹ãšå®å šæ§ïŒCompliance and SafetyïŒ: çµç¹ã®ããªã·ãŒãèŠå¶èŠä»¶ãã¢ããªã±ãŒã·ã§ã³åºæã®ã¬ã€ãã©ã€ã³ã«åŸãããããã¡ã€ã³ãã¥ãŒãã³ã°æžã¿ã¢ãã«ãåŠç¿ãããã
èšèªã»æåé©å¿ïŒLanguage or Cultural AdaptationïŒ: æ±çšã¢ãã«ã§ã¯äžååãªå Žåã«ããŒãããåŠç¿ããã³ã¹ãããããã«ãç¹å®ã®èšèªãæ¹èšãæåçæèã«åãããŠèšèªã¢ãã«ã調æŽããã
ãµããŒãããããã¡ã€ã³ãã¥ãŒãã³ã°ææ³
Supervised Finetuning
Direct Preference Optimization
Reinforcement Finetuning
æåž«ãããã¡ã€ã³ãã¥ãŒãã³ã°ïŒSFTïŒ
Supervised Fine-TuningïŒSFTïŒã¯ãç¹å®ã¿ã¹ã¯åãã«å ¥åã»åºåãã¢ã§äºååŠç¿æžã¿ã¢ãã«ãåŠç¿ããããåºç€çãªãã¬ãŒãã³ã°ææ³ã§ããããã«ãããèŠçŽã質åå¿çãã³ãŒãçæãªã©ã«ãããŠãå ã®ããŒã¹ã¢ãã«ã®ç¥èãä¿æãããŸãŸãããæ£ç¢ºã§äžè²«æ§ãããã¿ã¹ã¯ç¹åããå¿çãè¿ããããã«ãªããŸãã
SFT ã®æé©ãªãŠãŒã¹ã±ãŒã¹
ããã¹ãåé¡ãšã©ããªã³ã°
質åå¿çãšç¥èæœåº
ããã¹ãèŠçŽ
ã³ãŒãçæãšåæ
æ§é ååºåãšãã©ãŒããã
ãã¡ã€ã³åºæèšèªãŸãã¯ã¹ã¿ã€ã«ã®ã¢ã©ã€ã³ã¡ã³ã
ãã«ããšãŒãžã§ã³ããŸãã¯ããŒã«åŒã³åºãã¯ãŒã¯ãããŒ
ä»çµã¿
åºå®ãããäŸã®ã»ãããã¢ãã«ã«æäŸããäžããããå ¥åã«å¯ŸããŠæãŸããåºåãçæããããã«åŠç¿ãããŸããããã¯ãäŸããåŠã¶ïŒlearn by exampleïŒãã¢ãããŒãã§ãã

æåž«ãããã¡ã€ã³ãã¥ãŒãã³ã°ã®ã³ãŒãã¹ãããã
Microsoft Foundry SDK ãš Foundry UI ã§ãµããŒããããŠããŸãã以äžã¯ Microsoft Foundry SDK ã䜿ã£ãã³ãŒãã¹ããããã®äŸã§ãã
import os
from dotenv import load_dotenv
from azure.identity import DefaultAzureCredential
from azure.ai.projects import AIProjectClient
load_dotenv()
endpoint = os.environ.get("AZURE_AI_PROJECT_ENDPOINT")
model_name = os.environ.get("MODEL_NAME")
# Define dataset file paths
training_file_path = "training.jsonl"
validation_file_path = "validation.jsonl"
credential = DefaultAzureCredential()
project_client = AIProjectClient(endpoint=endpoint, credential=credential)
openai_client = project_client.get_openai_client()
with open(validation_file_path, "rb") as f:
validation_file = openai_client.files.create(file=f, purpose="fine-tune")
openai_client.files.wait_for_processing(validation_file.id)
with open(training_file_path, "rb") as f:
train_file = openai_client.files.create(file=f, purpose="fine-tune")
openai_client.files.wait_for_processing(train_file.id)
fine_tune_job = openai_client.fine_tuning.jobs.create(
model=model_name,
training_file=train_file.id,
validation_file=validation_file.id,
method={
"type": "supervised",
"supervised": {"hyperparameters": {"n_epochs": 3, "batch_size": 1, "learning_rate_multiplier": 1.0}},
},
suffix="pubmed-summarization"
)
ããŒã¿åœ¢åŒã®äŸ
åŠç¿ããŒã¿ã®ãµã³ãã«ã¯æå° 10 è¡ãå«ãå¿ èŠããããŸãã
{
"messages": [
{
"role": "system",
"content": "You are a medical research summarization assistant. Create concise, accurate abstracts of medical research articles that capture the key findings and methodology."
},
{ "role": "user", "content": "Summarize this medical research article:\n\n[full article text]" },
{ "role": "assistant", "content": "[generated abstract]" }
]
}
Cookbooks: PubMed å»åŠç ç©¶èŠçŽããŒã¿ã»ãããçšãã SFT
ãã® cookbookïŒfine-tuning/Demos/SFT_PubMed_Summarization at main · microsoft-foundry/fine-tuningïŒã¯ãAzure AI äžã§ PubMed Medical Research Summarization ããŒã¿ã»ãããçšã㊠Supervised Fine-TuningïŒSFTïŒã§èšèªã¢ãã«ããã¡ã€ã³ãã¥ãŒãã³ã°ããæ¹æ³ã瀺ããŸãã
cookbook ãå®è¡ããåŸãFoundry Portal ã«ç§»åããŠãžã§ã詳现ãç£èŠã§ããŸãã
Microsoft Foundry ã«ããããã¡ã€ã³ãã¥ãŒãã³ã°ãžã§ãç»é¢
Microsoft FoundryïŒhttps://ai.azure.comïŒã«ç§»åããfine-tuning ã»ã¯ã·ã§ã³ã§ãžã§ã詳现ãšå®è¡é²æã確èªããŸãã

äž»ãªãã€ã©ã€ã
ãã® cookbook ã§ã¯ãGPT4.1 ãããŒã¹ã¢ãã«ãšããŠäœ¿çšããKaggle ã® PubMed Article Summarization Dataset ãåç §ãã¬ãŒãã³ã°ããŒã¿ã»ãããšããŠäœ¿çšããŸãã
åææ¡ä»¶: Azure ãµãã¹ã¯ãªãã·ã§ã³ãš Microsoft Foundry ãããžã§ã¯ããããããšãAzure AI User ããŒã«ãå¿ èŠã§ãå¿ èŠãªã¢ãã«ã«ã¢ã¯ã»ã¹ã§ããAI Foundry ãããžã§ã¯ããã»ããã¢ããæžã¿ã§ããããšã
ããŒã¿ã»ããæºå: ãã¡ã€ã³ãã¥ãŒãã³ã°ã§ã¯ããŒã¿ã»ããã JSONL 圢åŒã§ããããšãæåŸ ãããŸããJSONL ã¯ããã§å ¥æã§ããŸã: ãã® Cookbook ãµã³ãã«ã§å©çšå¯èœãª training.jsonl ãš validation.jsonl ã䜿çšããŠãã ããã
ãã¡ã€ã³ãã¥ãŒãã³ã°ãžã§ã: æ¢å®ã®ãã€ããŒãã©ã¡ãŒã¿ã§æ§æãããã¡ã€ã³ãã¥ãŒãã³ã°ãžã§ããå®è¡ããŸãã
ãããã€: ä»»æã§ããã¡ã€ã³ãã¥ãŒãã³ã°æžã¿ã¢ãã«ããµãŒããŒã¬ã¹ãšã³ããã€ã³ããžãããã€ãããµã³ãã«æšè«ãå®è¡ããŸãã
