2026年版 Hugging Face Transformers完全ガイド：NLP、ファインチューニング、面接対策

# serve_model.py
# Start an OpenAI-compatible inference server from the command line
# transformers serve --model meta-llama/Llama-4-Scout-17B-16E-Instruct --compile

# Or use the Python API directly
from transformers import pipeline

# The pipeline API remains the fastest way to get predictions
classifier = pipeline("sentiment-analysis", model="distilbert-base-uncased-finetuned-sst-2-english")
results = classifier(["Transformers v5 simplifies everything.", "Legacy code migration is painful."])
print(results)
# [{'label': 'POSITIVE', 'score': 0.9998}, {'label': 'NEGATIVE', 'score': 0.9994}]

# load_model.py
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load tokenizer and model — architecture detected automatically
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=3)

# Tokenize input text with padding and truncation
inputs = tokenizer(
    "Hugging Face makes NLP accessible.",
    return_tensors="pt",     # Return PyTorch tensors
    padding=True,
    truncation=True,
    max_length=128
)

# Run inference with no gradient computation
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.softmax(outputs.logits, dim=-1)
    print(f"Class probabilities: {predictions}")

# tokenization_demo.py
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
text = "Transformers handle tokenization automatically."

# Step-by-step tokenization
tokens = tokenizer.tokenize(text)          # Split into subwords
print(f"Tokens: {tokens}")
# ['transformers', 'handle', 'token', '##ization', 'automatically', '.']

ids = tokenizer.convert_tokens_to_ids(tokens)  # Convert to numeric IDs
print(f"IDs: {ids}")
# [19081, 5765, 19204, 6032, 8073, 1012]

# The encode method does both steps plus adds special tokens
encoded = tokenizer.encode(text, add_special_tokens=True)
print(f"Encoded with special tokens: {encoded}")
# [101, 19081, 5765, 19204, 6032, 8073, 1012, 102]
# 101 = [CLS], 102 = [SEP]

# finetune_lora.py
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments, Trainer
from peft import LoraConfig, get_peft_model, TaskType
from datasets import load_dataset

# Load base model and tokenizer
model_name = "Qwen/Qwen3-0.6B"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto")

# Configure LoRA — only 0.5-2% of parameters become trainable
lora_config = LoraConfig(
    task_type=TaskType.CAUSAL_LM,
    r=16,                   # Rank of the low-rank matrices
    lora_alpha=32,          # Scaling factor
    lora_dropout=0.05,      # Dropout for regularization
    target_modules=["q_proj", "v_proj"],  # Which attention layers to adapt
)

# Wrap the model with LoRA adapters
model = get_peft_model(model, lora_config)
model.print_trainable_parameters()
# trainable params: 1,572,864 || all params: 631,000,000 || trainable%: 0.25

# Load and tokenize dataset
dataset = load_dataset("tatsu-lab/alpaca", split="train[:5000]")

def tokenize(example):
    return tokenizer(example["text"], truncation=True, max_length=512, padding="max_length")

tokenized = dataset.map(tokenize, batched=True, remove_columns=dataset.column_names)

# Configure training
training_args = TrainingArguments(
    output_dir="./lora-qwen",
    num_train_epochs=3,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,  # Effective batch size = 16
    learning_rate=2e-4,
    bf16=True,                      # Use bfloat16 mixed precision
    logging_steps=50,
    save_strategy="epoch",
)

# Train
trainer = Trainer(model=model, args=training_args, train_dataset=tokenized)
trainer.train()

# quantize_model.py
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
import torch

# Configure 4-bit quantization
quant_config = BitsAndBytesConfig(
    load_in_4bit=True,                      # Use 4-bit precision
    bnb_4bit_compute_dtype=torch.bfloat16,  # Compute in bfloat16
    bnb_4bit_quant_type="nf4",              # NormalFloat4 quantization
    bnb_4bit_use_double_quant=True,         # Quantize the quantization constants
)

# Load quantized model — fits in ~4GB VRAM instead of ~14GB
model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-7B-Instruct-v0.3",
    quantization_config=quant_config,
    device_map="auto",  # Automatically distribute across available GPUs
)

tokenizer = AutoTokenizer.from_pretrained("mistralai/Mistral-7B-Instruct-v0.3")

# Inference works identically to the non-quantized model
inputs = tokenizer("Explain quantization in one sentence:", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=50)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))