7.1.4 Pretrained Language Models at a Glance

One Practical Judgment

A pretrained model is not a magic model that knows your business. It is a reusable language foundation. Your job is to decide the cheapest reliable way to adapt that foundation to the task.

The Mental Model

Before pretraining became common, every NLP task often needed its own model and data pipeline. Modern NLP starts differently:

large general corpus→pretrained foundation→task adaptation→product behavior

The foundation has already learned useful language patterns. Your task usually needs one of these adaptations:

• prompt the model better;
• retrieve missing knowledge with RAG;
• train a small task head;
• fine-tune with LoRA or full updates;
• evaluate and guardrail the behavior.

What Pretraining Gives You

Pretraining usually gives three practical assets:

Asset	What it means	Example use
reusable representations	text already maps to useful hidden states	classification, ranking, retrieval
reusable generation ability	the model can continue or transform text	chat, writing, code generation
reusable language priors	grammar, common patterns, frequent facts	fewer examples needed downstream

It does not guarantee current knowledge, correct business policy, or safe behavior. Those still need data, retrieval, evaluation, and deployment controls.

Lab: Shared Foundation + Two Task Heads

This toy example does not train a real LLM. It shows the structure: one shared encoder, two different heads.

from math import exp

word_vectors = {
    "refund": [0.9, 0.8, 0.1],
    "order": [0.8, 0.7, 0.2],
    "password": [0.1, 0.2, 0.9],
    "reset": [0.1, 0.1, 0.95],
    "great": [0.7, 0.2, 0.1],
    "bad": [0.2, 0.8, 0.1],
}


def encode(text):
    tokens = text.lower().split()
    valid = [word_vectors[token] for token in tokens if token in word_vectors]
    dim = len(valid[0])
    return [sum(vec[i] for vec in valid) / len(valid) for i in range(dim)]


def dot(a, b):
    return sum(x * y for x, y in zip(a, b))


def softmax(scores):
    exps = [exp(score) for score in scores]
    total = sum(exps)
    return [value / total for value in exps]


intent_head = {
    "refund_intent": [1.0, 0.9, 0.1],
    "password_intent": [0.1, 0.2, 1.0],
}

sentiment_head = {
    "positive": [1.0, 0.2, 0.0],
    "negative": [0.1, 1.0, 0.0],
}


def classify(vector, head):
    labels = list(head.keys())
    scores = [dot(vector, head[label]) for label in labels]
    probs = softmax(scores)
    best = max(zip(labels, probs), key=lambda item: item[1])
    rounded = dict(zip(labels, [round(prob, 3) for prob in probs]))
    return best, rounded


for text in ["refund order", "reset password"]:
    vector = encode(text)
    best, probs = classify(vector, intent_head)
    print("intent:", text, "->", best, probs)

for text in ["great refund", "bad refund"]:
    vector = encode(text)
    best, probs = classify(vector, sentiment_head)
    print("sentiment:", text, "->", best, probs)

Expected output:

intent: refund order -> ('refund_intent', 0.7604230019887309) {'refund_intent': 0.76, 'password_intent': 0.24}
intent: reset password -> ('password_intent', 0.654188113761243) {'refund_intent': 0.346, 'password_intent': 0.654}
sentiment: great refund -> ('positive', 0.5793242521487495) {'positive': 0.579, 'negative': 0.421}
sentiment: bad refund -> ('negative', 0.5361866202317948) {'positive': 0.464, 'negative': 0.536}

Read it like this:

• encode() is the shared foundation.
• intent_head and sentiment_head are task-specific adapters.
• The foundation is reused; only the final decision layer changes.
• Real models do this with millions or billions of learned parameters instead of hand-written vectors.

Main Model Families

Family	Typical mask / flow	Strong at	Examples
Encoder-only	reads input bidirectionally	classification, extraction, matching, embeddings	BERT-style models
Decoder-only	predicts next token causally	chat, completion, code, tool use	GPT/LLaMA/Qwen-style models
Encoder-decoder	reads input, then generates output	translation, summarization, structured generation	T5/BART-style models

Use this as a first filter, not a final rule. Modern systems often combine families with retrieval, tools, and serving constraints.

Choose an Adaptation Path

Situation	Usually try first	Why
model already knows the task format	prompt improvement	fastest iteration
answer depends on private or fresh knowledge	RAG	update knowledge without changing weights
you need a stable label or score	task head / classifier	cheaper and easier to evaluate
style or domain behavior must shift	LoRA / PEFT	changes behavior with manageable cost
task is deeply specialized and data is strong	full fine-tuning	maximum flexibility, highest risk/cost

The decision is engineering, not ideology. Choose the smallest change that passes evaluation.

Common Failure Modes

• Pretraining data mismatch: the model learned broad language, not your exact policy.
• Stale knowledge: the model may not know recent facts.
• Contamination: benchmark or test data may appear in training-like corpora.
• Over-adaptation: fine-tuning can improve one behavior while damaging others.
• Evaluation gaps: a demo prompt can look good while edge cases fail.

Evidence to Keep

Keep this page’s proof of learning as a small evidence card:

Foundation

what the pretrained model already knows

Task Head

which part is task-specific

Adaptation Path

prompt, feature use, fine-tune, or PEFT

Eval Case

one example that proves transfer worked or failed

Risk

pretrained behavior is broad, not automatically aligned to your task

Exercises

1. Add a topic_head to the lab with labels account_topic and commerce_topic.
2. Change the vector for bad. How does sentiment confidence change?
3. For a support bot with private policies, would you start with prompt, RAG, task head, or fine-tuning? Explain.
4. List two checks you would run before trusting a pretrained model in production.
5. Explain why “bigger model” and “better task fit” are not the same thing.

Reference implementation and walkthrough

1. A topic_head would reuse the pretrained representation and map it to task labels such as account_topic and commerce_topic. The head is task-specific; the foundation remains reusable.
2. Changing the bad vector changes how strongly the sentiment head sees negative evidence. Confidence may drop, flip, or become less stable depending on where the vector moves.
3. For private policies, start with RAG in most cases because the knowledge is private, changing, and needs citations. Fine-tuning is worth considering later for stable behavior patterns or repeated formatting needs.
4. Two minimum checks are task evaluation on representative data and failure review on sensitive cases. Privacy, latency, cost, and bias checks are also production requirements.
5. A bigger model may have broader capability, but task fit depends on data, instructions, retrieval, evaluation, and operational constraints. Size is only one input to the decision.

Summary

Pretraining changes the workflow:

do not relearn language every time→reuse a foundation→adapt with evidence

Once you see that pattern, prompt engineering, RAG, fine-tuning, alignment, and agents all become different ways to steer the same reusable foundation.