ModernCS
Join for free
Hugging Face

Intelligence · Taught with Hugging Face

LLM / Fine-tuning

Prompting, retrieval, agents and fine-tuning. You measure a model before changing it, ground its answers in your own data, and put guardrails on it before anyone else uses it.

Download the course plan
  • Intermediate
  • 5 phases
  • 23 sessions
  • First session free

Syllabus

5 phases · 23 sessions · each phase ends in something you have built

1

Driving the Model

Get output you can reproduce instead of a demo that worked once.

$72

this phase

  1. 1.1

    Tokens, Sampling & Context

    Free preview

    What the model actually predicts, and what temperature, top-p, and a filling context window do to it.

    Open
  2. 1.2

    Chat Templates

    Apply the tokenizer's own template, then watch a mismatched one quietly wreck output quality.

  3. 1.3

    Structured Output

    JSON schemas, constrained decoding, and validation that fails loudly instead of silently.

  4. 1.4

    Prompting Reasoning Models

    Decomposition, few-shot, and why chain-of-thought prompts now hurt a model that already reasons.

By the end of this phase:A structured-output task that returns schema-valid JSON on fifty consecutive runs.
2

Evals Before Anything Else

Build the measurement you will judge every later change against.

$72

this phase

  1. 2.1

    Collecting Real Failures

    Turn logged mistakes into a labeled eval set instead of inventing test cases at your desk.

  2. 2.2

    Graders That Can't Argue

    Exact match, schema checks, unit tests, and reshaping a task so a grader can verify it.

  3. 2.3

    LLM as Judge, Honestly

    Position, verbosity, and self-preference bias, and checking your judge against human labels.

  4. 2.4

    Benchmarks & Contamination

    Run LightEval, read a leaderboard, and work out whether a score was memorized.

By the end of this phase:An eval harness of forty cases drawn from real failures, with a locked baseline score.
3

Grounding & Context

Give the model your documents and prove retrieval earns what it costs.

$72

this phase

  1. 3.1

    Embeddings & Vector Search

    Turn meaning into geometry, and measure recall@k before blaming the model.

  2. 3.2

    Retrieval That Holds Up

    Chunking, hybrid keyword plus dense search, reranking, and citations a reader can check.

  3. 3.3

    Long Context vs Retrieval

    Run both on your eval, then compare accuracy, latency, and cost per query before you choose.

  4. 3.4

    Context Engineering

    Token budgets, compaction, and why answers degrade long before the window is full.

By the end of this phase:A cited assistant benchmarked against a long-context baseline on the same eval set.
4

Agents, Tools & Rails

Let a model take real actions without handing an attacker your systems.

$72

this phase

  1. 4.1

    Tool Calling

    Write tool schemas a model can actually use, and handle the call it gets wrong.

  2. 4.2

    MCP & the Protocol Layer

    Connect to MCP servers, publish one of your own, and see where agent-to-agent protocols sit.

  3. 4.3

    Code Agents & Sandboxes

    Let the model write Python for its whole plan, then run it where it can't hurt you.

  4. 4.4

    Tracing a Failed Run

    Read a full trace, classify the failure, and fix the step that actually broke.

  5. 4.5

    Injection & the Four Rails

    Guard input, output, retrieved text, and tool arguments, then red-team your own agent.

By the end of this phase:A sandboxed agent wired to MCP servers, traced end to end, with input and tool-call rails.
5

Making the Model Yours

Change a model's behavior and prove on your own eval that you improved it.

$72

this phase

  1. 5.1

    Should You Fine-Tune?

    Rule out prompting and retrieval first, and spot the knowledge problem tuning will never fix.

  2. 5.2

    Building the Training Set

    Format the data, mask the prompt out of the loss, and audit what you're about to train on.

  3. 5.3

    Adapters Past Plain LoRA

    LoRA, rank-stabilized LoRA, and DoRA compared on your task under one memory budget.

  4. 5.4

    Quantized vs Full-Precision Tuning

    QLoRA on one consumer GPU, and the model families where 4-bit training now costs you accuracy.

  5. 5.5

    Preference & Reward Tuning

    DPO on preference pairs, then GRPO with a verifiable reward once a grader can score it.

  6. 5.6

    Serving It & Signing It

    Quantized inference on vLLM, a measured cost per request, and a model card stating the limits.

By the end of this phase:A published adapter that beats the base model on your eval, served with a model card.

Tools you will use

  • Hugging Face Transformers
  • TRL
  • PEFT
  • Unsloth
  • Sentence Transformers
  • vLLM
  • smolagents / MCP
  • LightEval

What you will build

  • The Eval Harness

    Forty real failures, graders, and a baseline to beat

  • Grounded Agent

    MCP tools, citations, sandbox, injection rails

  • Your Own Adapter

    One GPU, measured gain, published with a model card

LLM / Fine-tuning · ModernCS