Skip to content

Module C · Topic 02 — LLM Fundamentals

CLO-3 · Maps to: Session 25 · Notebook 15 · Lab 25 · Level: Beginner

1. Beginner explanation

A Large Language Model (LLM) is a program that predicts the next word. Show it "The capital of France is", and it completes: "Paris". That's the entire trick — repeated billions of times, with a giant neural network trained on an enormous amount of text. It has no database of facts and no ability to compute. It has patterns. When you ask it something, it is not looking anything up; it is generating the most plausible-sounding continuation. That's why it can sound brilliant and be wrong with the same confidence — a behaviour called hallucination.

2. Conceptual explanation (the WHY)

Internally, text is split into tokens (roughly word pieces: "data" might be one token, "sci" + "ence" two). The model assigns a probability to every possible next token based on the context so far — a softmax over its whole vocabulary. Generating text = sampling from those probabilities, one token at a time.

Three knobs matter for data work:

  • Temperature (0–1+): low → always pick the most probable token (predictable, dry); high → pick unlikely tokens (creative, sloppy). For data analysis: low (0–0.3).
  • System prompt: instructions that set the role and rules for the whole conversation, separate from the user's message.
  • Context window: how many tokens the model can "see" at once. Whole dataframes do not fit — that's why you send schema and samples, not rows.

The model's knowledge is frozen at training time and stored as weights, not retrievable facts. It cannot count reliably, cannot do exact arithmetic, and will confidently answer questions whose answer it never learned. Anything factual or numerical it says must be verified.

3. Simple diagram

 "The capital of       tokens: [The][ capital][ of]
  France is"
        │
        ▼
  ┌───────────────────────────┐
  │ Neural network (weights)  │   trained once on billions of words
  └────────────┬──────────────┘
               │ assigns probability to EVERY next token
               ▼
   "Paris"   0.82  ← most likely → chosen (low temperature)
   "a"       0.07
   "the"     0.04
   "Berlin"  0.01  ← plausible-sounding, WRONG, but possible!
   ...        ...
        │
        ▼
   Output: "Paris"   -- and the process repeats for the next token

4. Python examples

import importlib

def module_available(name):
    try:
        importlib.import_module(name)
        return True
    except ImportError:
        return False

if module_available("ollama"):
    import ollama

    try:
        # system message sets rules; user message asks the task
        messages = [
            {"role": "system",
             "content": "You are a careful data assistant. If you are not "
                        "sure, say 'I am not sure' instead of guessing."},
            {"role": "user",
             "content": "Compute 123 * 456 exactly."},
        ]
        reply = ollama.chat(model="llama3.2", messages=messages)
        print("MODEL SAYS:", reply["message"]["content"])
        # Always compare against ground truth:
        print("PYTHON SAYS:", 123 * 456)
    except Exception as e:
        print("Ollama not running — start it and pull llama3.2:", e)
else:
    print("ollama package not installed:  pip install ollama")

The point of the example: the model may or may not get the arithmetic right. Never let a model be your calculator.

5. Practical exercise (20 min)

  1. In a notebook, ask ollama.chat the same arithmetic question 3 times and compare answers (LLMs are not deterministic — sampling!).
  2. Ask a factual question about this course's schedule: "When is the midterm?" — the model has never seen your syllabus. Observe what it invents. This is the cleanest hallucination demo possible.
  3. Repeat with a system prompt "If unsure, say you don't know" and note the difference in behaviour (not a guarantee — just better behaviour).
  4. Write one paragraph: where can an LLM help your data workflow, and where must it never be the final authority?

6. Common errors

Error Why it happens Fix
"The model got my facts wrong" treating it as a fact database it predicts text; verify facts externally
"The answer changes every run" sampling is random set temperature low; treat output as draft
"It can't do my arithmetic" not a calculator compute with Python; use the model only for phrasing/planning
"It made up a column name" hallucinated schema give it real column names in the prompt
"It ignored my instructions" vague or buried system prompt make the system prompt short, specific, first

7. Limitations

  • Hallucination: fluent, confident, wrong. No built-in "I checked that".
  • No exact computation: arithmetic, counting, and comparisons are unreliable.
  • Stale knowledge: training data has a cutoff; course-specific facts are absent.
  • Context limits: cannot "read" whole datasets; only small excerpts fit.
  • No source memory: it does not remember where it learned something, so it cannot cite its own claims.

8. Responsible AI considerations

  • Verification culture: every model claim about data is a hypothesis, checked against actual data before use.
  • Privacy: anything you paste into a cloud LLM leaves your machine. Course policy: local Ollama for real data; never paste grades, IDs, or personal data.
  • Honest disclosure: the report/README lists every LLM-assisted step (tool, version, prompt, how output was verified).
  • Confidence is not accuracy: teach students to distrust fluent certainty — the model's confidence has no relationship to correctness.

9. Assessment questions

Q1. In one sentence, what does an LLM actually do when you ask it a question? CLO-3 · Understand · Easy Answer: It predicts the most probable next token, repeatedly — it does not look up facts or compute.

Q2. Why is an LLM's confident answer about your course's exam date likely wrong? CLO-3 · Analyze · Medium Answer: The model never saw your syllabus; its knowledge is frozen at training time, so anything course-specific is generated (hallucinated) text that only sounds plausible.

Q3. Your analysis must be reproducible. A teammate re-ran the LLM step and got a different answer. What caused it and what do you do? (CLO-3 · Evaluate · Medium) Answer: Sampling is non-deterministic. Fix temperature (low), save the prompt + model + version + output in the audit log, and verify the answer against data by hand so the result doesn't depend on a lucky sample.

Q4. Give one task an LLM is genuinely good at in a data workflow and one task it must never do alone. CLO-3 · Evaluate · Medium Answer: Good: translating a question into a first draft of pandas code, explaining an error, drafting a report section. Never alone: computing final numbers or making a decision — those must be verified by code/data.