メニュー
AI and ML expert covering PyTorch, TensorFlow, Hugging Face, scikit-learn, LLM integration, RAG pipelines, MLOps, and production ML systems
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/oimiragieo/agent-studio --skill ai-ml-expertコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
SOC 職業分類に基づく
SKILL.md を表示中
| name | ai-ml-expert |
| description | AI and ML expert covering PyTorch, TensorFlow, Hugging Face, scikit-learn, LLM integration, RAG pipelines, MLOps, and production ML systems |
| version | 2.1.0 |
| model | sonnet |
| invoked_by | both |
| user_invocable | true |
| tools | ["Read","Write","Edit","Bash","Grep","Glob","WebSearch"] |
| best_practices | ["Reproducibility first — fix random seeds, log all hyperparameters","Data quality and preprocessing as the foundation of every model","Evaluate with multiple metrics aligned to business goals","Test data never seen during training (rigorous splits)","Prefer fine-tuning and transfer learning over training from scratch"] |
| error_handling | graceful |
| streaming | supported |
| verified | true |
| lastVerifiedAt | "2026-02-19T00:00:00.000Z" |
| source | builtin |
| trust_score | 100 |
| provenance_sha | 54c7d87f033bd4e4 |
When reviewing or writing PyTorch code, apply these guidelines:
torch.nn.Module for all model definitions; avoid raw function-based modelsmodel.to(device), tensor.to(device)model.train() and model.eval() context switches appropriatelyoptimizer.zero_grad() at the top of the training looptorch.no_grad() or @torch.inference_mode() for all inference codepin_memory=True) and use multiple workers in DataLoader for GPU trainingtorch.compile() (PyTorch 2.x) for production inference speedupsF.cross_entropy over manual softmax + NLLLoss (numerically stable)When reviewing or writing TensorFlow code, apply these guidelines:
tf.data.Dataset pipelines over manual batching for scalabilitytf.function for graph execution on performance-critical pathstf.keras.mixed_precision.set_global_policy('mixed_float16')tf.saved_model for portable model export; avoid picklingWhen reviewing or writing Hugging Face code, apply these guidelines:
padding=True and truncation=True when tokenizing batchesAutoModel, AutoTokenizer, and AutoConfig for checkpoint portabilitymodel.gradient_checkpointing_enable() to reduce memory for large modelsTrainer API for standard fine-tuning; use custom loops only when Trainer is insufficientTRANSFORMERS_CACHE environment variable in CI/CD pipelinesWhen reviewing or writing scikit-learn code, apply these guidelines:
Pipeline to chain preprocessing and model steps; prevents data leakageStratifiedKFold for classification tasks with class imbalanceGridSearchCV or RandomizedSearchCV for hyperparameter tuning.fit() only on training data; transform test data with the fitted transformerjoblib.dump / joblib.load (faster than pickle for large arrays)"Think step by step...") for complex reasoning taskstemperature=0 for deterministic, fact-based outputs; increase for creative tasks# Standard RAG pipeline components
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.vectorstores import FAISS # or Chroma, Pinecone, Weaviate
from langchain.chains import RetrievalQA
# 1. Embed and index documents
embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")
vectorstore = FAISS.from_documents(documents, embeddings)
# 2. Retrieve relevant chunks
retriever = vectorstore.as_retriever(search_kwargs={"k": 4})
# 3. Generate with retrieved context
chain = RetrievalQA.from_chain_type(llm=llm, retriever=retriever)
RAG best practices:
LCEL (LangChain Expression Language) for composable chainsRunnableParallel for concurrent retrieval stepsLangGraph for stateful multi-agent workflows with cyclesRunnableRetry for unreliable external calls# Standard PyTorch training loop with best practices
for epoch in range(num_epochs):
model.train()
for batch in train_dataloader:
optimizer.zero_grad()
inputs, labels = batch["input_ids"].to(device), batch["labels"].to(device)
outputs = model(inputs)
loss = criterion(outputs, labels)
loss.backward()
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0) # gradient clipping
optimizer.step()
scheduler.step()
# Validation loop
model.eval()
with torch.no_grad():
for batch in val_dataloader:
# evaluate...
Key standards:
max_norm=1.0) for stability in Transformer trainingsmoothing=0.1) to reduce overconfidencefrom peft import LoraConfig, get_peft_model, TaskType
lora_config = LoraConfig(
task_type=TaskType.CAUSAL_LM,
r=16, # LoRA rank
lora_alpha=32, # scaling factor
target_modules=["q_proj", "v_proj"],
lora_dropout=0.05,
)
model = get_peft_model(base_model, lora_config)
model.print_trainable_parameters() # verify < 1% parameters trainable
PEFT guidelines:
r=8 to r=64; higher rank = more capacity, more memoryimport mlflow
with mlflow.start_run():
mlflow.log_params({"learning_rate": lr, "batch_size": bs, "epochs": epochs})
mlflow.log_metrics({"train_loss": loss, "val_accuracy": acc}, step=epoch)
mlflow.pytorch.log_model(model, "model")
import wandb
wandb.init(project="my-project", config={"lr": 1e-4, "epochs": 10})
wandb.log({"train_loss": loss, "val_f1": f1_score})
wandb.finish()
MLOps standards:
torch.manual_seed(42), np.random.seed(42), random.seed(42)| Task | Primary Metrics | Secondary Metrics |
|---|---|---|
| Binary Classification | AUC-ROC, F1, Precision/Recall | Calibration (Brier Score) |
| Multi-class | Macro F1, Weighted F1, Cohen's Kappa | Confusion Matrix |
| Regression | RMSE, MAE, R² | Residual Analysis |
| NLP Generation | BLEU, ROUGE, BERTScore | Human Evaluation |
| Ranking/Retrieval | NDCG@k, MRR, MAP | Hit Rate@k |
| LLM Evaluation | LLM-as-judge, exact match, pass@k | Hallucination Rate |
torch.onnx.export(model, ...)torch.quantization.quantize_dynamic(model, ...)# Example: data drift detection with Evidently
from evidently.report import Report
from evidently.metric_preset import DataDriftPreset
report = Report(metrics=[DataDriftPreset()])
report.run(reference_data=reference_df, current_data=production_df)
report.save_html("drift_report.html")
Monitoring standards:
# Proper train/test split to avoid leakage
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=42, stratify=y # stratify for classification
)
# Fit scaler ONLY on training data
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test) # transform only, never fit_transform
Standards:
torch.manual_seed(42), np.random.seed(42), random.seed(42) and log via MLflow/W&B..transform() test; fitting on test causes data leakage and inflated performance estimates.| Anti-Pattern | Problem | Fix |
|---|---|---|
| Ignoring class imbalance | Model biased to majority class | Stratified sampling, class weights, SMOTE |
| No validation set | Overfitting undetected | Hold out 10-20% for validation |
| Optimizing a single metric | Missing failure modes | Multiple metrics (precision, recall, F1, AUC) |
| No baseline comparison | Cannot assess model quality | Establish heuristic baseline before ML |
| Accuracy on imbalanced data | Misleading performance estimate | Use F1, precision-recall curve, ROC-AUC |
| Data leakage (test in train) | Inflated performance estimates | Fit on train only; transform test with fitted obj |
| No error analysis | Cannot improve strategically | Analyze failure cases by error type |
| Training without checkpoints | Lost progress on failure | Save best model by validation metric |
| Mutable global random state | Non-reproducible experiments | Fix all seeds; log in experiment metadata |
| Embedding model in application | Cannot update model independently | Serve model via API (REST, gRPC) |
| No latency budget | Inference too slow for production | Profile and set SLO before deployment |
Training a Transformer classifier:
from transformers import AutoTokenizer, AutoModelForSequenceClassification, Trainer, TrainingArguments
tokenizer = AutoTokenizer.from_pretrained("distilbert-base-uncased")
model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased", num_labels=3)
def tokenize(batch):
return tokenizer(batch["text"], padding=True, truncation=True, max_length=512)
dataset = dataset.map(tokenize, batched=True)
training_args = TrainingArguments(
output_dir="./results",
num_train_epochs=3,
per_device_train_batch_size=16,
evaluation_strategy="epoch",
save_strategy="epoch",
load_best_model_at_end=True,
metric_for_best_model="f1",
)
trainer = Trainer(
model=model,
args=training_args,
train_dataset=dataset["train"],
eval_dataset=dataset["validation"],
compute_metrics=compute_metrics,
)
trainer.train()
Minimal RAG pipeline:
from langchain.embeddings import OpenAIEmbeddings
from langchain.vectorstores import Chroma
from langchain.chains import RetrievalQA
from langchain.chat_models import ChatOpenAI
vectorstore = Chroma.from_documents(docs, OpenAIEmbeddings())
retriever = vectorstore.as_retriever(search_kwargs={"k": 5})
qa = RetrievalQA.from_chain_type(ChatOpenAI(model="gpt-4o"), retriever=retriever)
answer = qa.run("What is the refund policy?")
This skill is used by:
developer — Implements ML models, data pipelines, and LLM integrationsresearcher — Investigates novel architectures and evaluates research papersarchitect — Designs ML system architecture and deployment topologysecurity-architect — Reviews data privacy, model security, and inference safetypython-backend-expert — NumPy, Pandas, async Python patternscode-analyzer — Static analysis and complexity metrics for ML codedebugging — Systematic debugging for training failures and inference errorsBefore starting:
cat .claude/context/memory/learnings.md
Check for:
After completing:
.claude/context/memory/learnings.md.claude/context/memory/issues.md.claude/context/memory/decisions.mdASSUME INTERRUPTION: Your context may reset. If it's not in memory, it didn't happen.