Course 3 — LLM Fine-Tuning Masterclass

Fine-Tuning
Masterclass

Course 1 said the model is 1.6% of an agent. This course zooms into that 1.6%. How do you steer, align, unalign, quantize, and deploy an LLM for local, sensitive-data, and calibrated-compliance use? The thesis: fine-tuning steers behavior; it does not teach knowledge. The model steers — the harness bounds. Eight pillars. Twenty-four modules. Ten deep-dives. Two capstones. One synthesis: uncensor the model so it executes; harness the model so it executes only what it should.

~30
hours
24
modules
2
capstones
10
deep-dives
1.6%

Course 1 said the model is 1.6% of an agent. Course 3 zooms in. Fine-tuning steers behavior; it does not teach knowledge. Every technique — SFT, DPO, GRPO, abliteration — redirects an already-capable base model. You are not teaching; you are steering. And steering is only safe inside a harness.

LayerWhat it isThis course
1. The BasePretrained weightsFT00–FT03
2. The AdapterLoRA/DoRA, <1% paramsFT08–FT09
3. The SteerSFT, DPO, GRPO, abliterationFT12–FT18
4. The ExportQuantize + serveFT19–FT20
5. The BoundaryThe harness (Courses 1 & 2A)FT21–FT23
Pillar 0 — Foundations
4 modules · ~3.5 hrs
FT00
The Steering Stack
The thesis. Fine-tuning steers behavior; it does not teach knowledge. The five-layer mental model.
60 MIN
FT01
VRAM Math: Can I Actually Run This?
The affordability gate. Three VRAM consumers, the method-vs-GPU decision, the three-question framework.
45 MIN
FT02
The Open Spectrum: Weights, Data, and Trust
Open-weights vs open-data vs open-recipe. The OSAID gap, the NTIA argument, why open-data wins for sensitive domains.
60 MIN
FT03
Base Model Selection
The five-dimension rubric. MiniCPM5-1B as the on-ramp; when to graduate to Qwen/Llama/DeepSeek.
45 MIN
Pillar 1 — Data
4 modules · ~4 hrs
FT04
Dataset Formats
The three canonical formats, apply_chat_template, SFT vs preference data.
45 MIN
FT05
Synthetic Data Generation
Magpie self-synthesis, teacher distillation, Evol-Instruct, distilabel.
75 MIN
FT06
Dedup, Filter, Decontaminate
MinHash, embedding dedup, perplexity filtering, decontamination. Where 80% of gains are.
60 MIN
FT07
Tokenizers & Chat Templates
The silent-bugs checklist. Cross-family templates, EOS, packing.
60 MIN
Pillar 2 — Parameter-Efficient Fine-Tuning
4 modules · ~4.5 hrs
FT08
LoRA & QLoRA
The low-rank decomposition, QLoRA's three innovations, the first real fine-tune.
75 MIN
FT09
DoRA, rsLoRA, and Modern PEFT
DoRA closes half the LoRA-to-full-FT gap. The 2026 sweet spot.
60 MIN
FT10
Full FT vs PEFT: The Decision
When to go full FT (rarely). The structural non-equivalence finding.
45 MIN
FT11
The Training Loop with TRL
TRL v1.0, the SFTTrainer, reading loss curves, the standard levers.
90 MIN
Pillar 3 — Alignment & Preferences
2 modules · ~2.25 hrs
FT12
SFT: The Baseline
Building the SFT mix, failure modes, the CPT sidebar.
60 MIN
FT13
The DPO Family and Preferences
DPO, IPO, KTO, ORPO, SimPO. The preference-method decision tree.
75 MIN
Pillar 4 — Reasoning Models
2 modules · ~2.75 hrs
FT14
GRPO and Verifiable Rewards
The flagship. Why RL for reasoning, GRPO mechanism, the R1/Qwen3 pipelines.
90 MIN
FT15
CoT Distillation and Rejection-Sampling FT
Distill reasoning from a strong teacher, RFT, the three-way decision.
75 MIN
Pillar 5 — Alignment Control (Uncensored)
3 modules · ~3.25 hrs
FT16
Why Uncensored: The Legitimate Use Cases
Over-refusal as operational defect. Security research, tool agents, medical/legal.
45 MIN
FT17
Abliteration: Refusal-Direction Orthogonalization
The Arditi finding, the diff-in-means pipeline, the capability-degradation trade-off.
90 MIN
FT18
Compliance via DPO and SFT
The higher-fidelity paths. Dolphin, Hermes. The three-way comparison.
60 MIN
Pillar 6 — Quantize & Deploy
2 modules · ~2.25 hrs
FT19
Quantization Formats
GGUF, AWQ, GPTQ, EXL2, MLX, FP8, FP4. The format decision matrix.
75 MIN
FT20
Serving Stacks
vLLM vs llama.cpp vs Ollama vs MLX. The no-telemetry/air-gap considerations.
60 MIN
Pillar 7 — Sensitive Domains
3 modules · ~3.25 hrs
FT21
HIPAA and BAA Elimination
Why local eliminates the BAA. The API coverage gaps.
60 MIN
FT22
Government, Military, and Air-Gapped Deployment
DoD IL2–IL7, the LoRA supply-chain risk, the defensive playbook.
75 MIN
FT23
Synthesis: Uncensor the Model, Harness the Model
THE bridge. The synthesis principle, the honest caveat, the full-stack picture.
60 MIN
Capstones
2 capstones · ~3.5 hrs
CAP1
The Air-Gapped Domain Model
A reproducible pipeline: open-data base → HIPAA-safe data → QLoRA → eval → GGUF → local serve.
120 MIN
CAP2
The Calibrated Uncensored Agent
Steer toward compliance, quantize, wrap in an eval'd harness. The bridge to Course 1.
90 MIN
Deep-Dives — 10 Fine-Tuning Studies
10 studies · live
FTDD-01
MiniCPM Family (OpenBMB)
The open-everything on-ramp hero.
FTDD-02
OLMo 2/3 + Tülu 3 (Ai2)
The fully-open comparison.
FTDD-03
Unsloth
Single-GPU speed + Dynamic 2.0 GGUF.
FTDD-04
TRL
The canonical post-training library.
FTDD-05
Axolotl
Config-driven multi-GPU production path.
FTDD-06
Dolphin / Hermes (Nous)
The uncensored lineages as engineering.
FTDD-07
DeepSeek-R1
Reasoning distillation reference.
FTDD-08
Qwen3
Hybrid thinking/non-thinking pipeline.
FTDD-09
llama.cpp + vLLM
Serving deep dive.
FTDD-10
distilabel
Synthetic data pipeline construction.

Prerequisite: Course 1 — the Master Course (or equivalent production harness experience). Course 3 assumes you already understand why the model is only 1.6% of an agent and why the harness is the other 98.4%. All twenty-four modules, both capstones, and all ten deep-dives are live now. Each module ships the same eight-artifact stack: teaching document, diagrams, slide deck, teaching script, flashcards, exam, lab specification, and web page. Every page is live on here.now for mobile review. Course 3 closes the loop opened in Course 1: the model steers, the harness bounds — and now you can build the steering.