A.9 Learning Resources Quick Reference

Use this page while building. Do not read it from top to bottom.

Environment checks

python --version
which python
pip --version
pip list
pwd
ls

For the docs site:

npm install
npm run start
npm run build

For NVIDIA GPU:

nvidia-smi

Baseline first

Task	Try first
Tabular classification/regression	Linear model or tree model
Text classification	TF-IDF + LogisticRegression
Image classification	Transfer learning
Named entity recognition	Rules/dictionary baseline, then sequence model
Document Q&A	Keyword/BM25 retrieval, then RAG
Agent tool use	Single Agent + one safe tool

Metrics

Task	First metrics
Balanced classification	Accuracy, F1
Imbalanced classification	Precision, Recall, F1, confusion matrix
Regression	MAE, RMSE, residual review
Retrieval / RAG	Hit@K, MRR, citation accuracy, human review
Agent	success rate, tool errors, cost, trace review

Training warning signs

Signal	Check first
Loss does not decrease	labels, loss function, learning rate, input format
Train good, validation poor	overfitting, leakage, distribution mismatch
Accuracy unchanged	weak features, wrong labels, model not learning
GPU out of memory	batch size, input length, model size
Unstable results	random seed, small data, inconsistent split

RAG checklist

1. Documents split correctly?
2. Retrieval returns the right chunks?
3. Answer includes sources?
4. Answer truly uses the retrieved content?
5. Permission filtering and no-answer behavior exist?

Agent checklist

1. Start with single-turn Q&A.
2. Add one tool.
3. Add strict parameter schema.
4. Add logs and trace replay.
5. Add permission boundary and stop condition.

Prompt template

You are a ____.
Your task is ____.
Input:
Output format:
Constraints:
If information is insufficient, say so clearly.

Minimal training loop

data = [(1.0, 2.0), (2.0, 4.0), (3.0, 6.0)]
w = 0.0
lr = 0.01

for epoch in range(3):
    total_loss = 0.0
    for x, y in data:
        pred = w * x
        error = pred - y
        total_loss += error * error
        grad = 2 * error * x
        w -= lr * grad
    print(f"epoch={epoch} w={w:.3f} loss={total_loss:.3f}")

Expected output:

epoch=0 w=0.521 loss=48.630
epoch=1 w=0.907 loss=26.580
epoch=2 w=1.192 loss=14.528

Read it as: data -> prediction -> loss -> gradients -> parameter update.