Fine-Tune Quest

Tracks

1. 01🎛️Fine-Tuning Decision

   0/7 lessons

   When to train, when not to

   Fine-tuning earns its keep only when prompt engineering and RAG hit a measurable ceiling. This track teaches you to find that ceiling before spending a dollar on training compute.

   Lesson list (7)

   • 01
     The Customization Spectrum~22 min · decision, prompt-engineering, rag
   • 02
     When Fine-Tuning Wins~20 min · use-cases, distillation, format, style
   • 03
     When Fine-Tuning Loses~18 min · anti-patterns, catastrophic-forgetting, rag
   • 04
     What Fine-Tuning Actually Does~20 min · gradients, transfer-learning, mental-model
   • 05
     Data Requirements & Cost Landscape~22 min · budget, dataset-size, infra, apple-silicon
   • 06
     Base vs Instruct Models~18 min · base-model, instruct, starting-point, chat-template
   • 07
     The Decision Flowchart & License Reality~22 min · decision-tree, licensing, evaluation, legal

   ✓Quiz · 5 questions→

2. 02📊Training Data

   0/5 lessons

   Garbage in, garbage out — at tenfold strength

   The single most impactful lever in any fine-tuning project. Hyperparameters move the needle by single digits; data quality moves it by tens. This track covers JSONL, formats, synthesis, cleaning, splits, and annotation.

   Lesson list (5)

   • 01
     Data Quality Beats Data Quantity~22 min · data-quality, curation, diversity
   • 02
     Data Formats: JSONL, OpenAI, Alpaca, ShareGPT~22 min · jsonl, openai-format, alpaca, sharegpt
   • 03
     Data Collection & Synthetic Generation~28 min · synthetic-data, distillation, production-logs
   • 04
     Data Cleaning, Validation & Splits~22 min · dedup, validation, splits, overfitting
   • 05
     Annotation Tools: Argilla, Label Studio, Scripts~20 min · annotation, argilla, label-studio, human-review

   ✓Quiz · 5 questions→

3. 03🔑OpenAI Fine-Tuning

   0/4 lessons

   The managed, zero-infrastructure path

   The fastest path from data to deployed fine-tuned model: upload JSONL, create a job, use the model. SFT, DPO, vision FT, and RFT — what each is for and when each is the right pick.

   Lesson list (4)

   • 01
     Supported Models & Methods~20 min · openai, sft, dpo, rft, pricing
   • 02
     Preparing & Uploading Data~22 min · openai, jsonl, upload, validation
   • 03
     Creating & Monitoring Jobs~22 min · openai, jobs, hyperparameters, monitoring
   • 04
     Using Fine-Tuned Models~20 min · openai, inference, evaluation, comparison

   ✓Quiz · 4 questions→

4. 04🧩PEFT and Adapters

   0/6 lessons

   Fine-tune smart, not hard

   Parameter-efficient fine-tuning — LoRA, QLoRA, DoRA, rsLoRA — updates 0.1–1% of parameters and gets ~95–99% of full-fine-tuning quality. The reason most modern projects fine-tune at all.

   Lesson list (6)

   • 01
     The Parameter Problem~18 min · full-fine-tuning, vram, memory
   • 02
     LoRA: The Breakthrough~22 min · lora, low-rank, decomposition
   • 03
     LoRA Hyperparameters~22 min · lora, rank, alpha, target-modules, dropout
   • 04
     QLoRA: Fine-Tune 70B on One GPU~24 min · qlora, quantization, nf4, double-quant, memory
   • 05
     LoRA Variants & Comparison~20 min · dora, lora-plus, rslora, adalora, vera
   • 06
     Merging & Multiple Adapters~22 min · merge, adapters, multi-adapter, apple-silicon

   ✓Quiz · 5 questions→

5. 05🤗Hugging Face Stack

   0/5 lessons

   The open-source path with TRL v1.0

   The full open-source training stack: transformers, datasets, peft, trl, accelerate. End-to-end QLoRA pipeline, push to Hub, convert to GGUF for Ollama. The path you take when you need full control.

   Lesson list (5)

   • 01
     The Hugging Face Stack~18 min · transformers, trl, peft, datasets, accelerate
   • 02
     Loading Models & Tokenizers~22 min · model-loading, quantization, chat-template
   • 03
     SFTTrainer: The High-Level API~24 min · trl, sft-trainer, sft-config, supervised-fine-tuning
   • 04
     Complete QLoRA Pipeline~28 min · qlora, end-to-end, wandb, pipeline
   • 05
     Saving, Pushing & Converting~22 min · hub, gguf, ollama, conversion, quantization

   ✓Quiz · 5 questions→

6. 06🏋️Training Technique

   0/6 lessons

   The knobs that make or break results

   Learning rate, batch size, loss curves, memory optimization, alignment (DPO/ORPO/SimPO), and evaluation. The hyperparameters and techniques that turn a working training run into a production-quality model.

   Lesson list (6)

   • 01
     Learning Rate & Scheduling~20 min · learning-rate, scheduler, cosine, warmup
   • 02
     Batch Size & Gradient Accumulation~18 min · batch-size, gradient-accumulation, effective-batch
   • 03
     Reading Loss Curves~20 min · loss-curve, overfitting, underfitting, early-stopping
   • 04
     Memory Optimization~22 min · bf16, gradient-checkpointing, flash-attention, memory
   • 05
     Alignment: DPO, ORPO, SimPO~22 min · dpo, orpo, simpo, alignment, preference
   • 06
     Evaluation Methods~20 min · evaluation, perplexity, llm-judge, metrics

   ✓Quiz · 5 questions→

7. 07📚Fine-Tune Projects

   0/5 lessons

   Real-world recipes from data to deployment

   End-to-end project recipes — instruction follower, domain expert, brand voice, synthetic data flywheels, and a top-10 list of common failure modes with debug fixes for each.

   Lesson list (5)

   • 01
     Project: Custom Instruction Follower~28 min · project, structured-output, json, instruction-follower
   • 02
     Project: Domain Expert~28 min · project, domain-expert, medical, legal, compliance
   • 03
     Project: Style Transfer & Brand Voice~24 min · project, style-transfer, brand-voice, marketing
   • 04
     Synthetic Data & Data Flywheels~24 min · synthetic, distillation, flywheel, ab-testing
   • 05
     Common Failures & Debugging~22 min · debugging, failures, gotchas, troubleshooting

   ✓Quiz · 5 questions→

8. 08🚀Deployment & Production

   0/5 lessons

   Serving fine-tuned models at scale

   Merge → convert → serve. Local with Ollama and llama.cpp. Cloud with vLLM and TGI. FastAPI wrappers, monitoring, multi-adapter S-LoRA, edge deployment. The full path from trained adapter to production endpoint.

   Lesson list (5)

   • 01
     Merging & Export~18 min · merge, export, deployment-pipeline
   • 02
     Local Deployment: Ollama & llama.cpp~24 min · ollama, llama-cpp, gguf, local, apple-silicon
   • 03
     Cloud Serving: vLLM & TGI~22 min · vllm, tgi, paged-attention, openai-compatible
   • 04
     API Wrapping & Monitoring~20 min · fastapi, monitoring, drift, sli
   • 05
     Edge Deployment & Multi-Adapter Patterns~20 min · edge, s-lora, multi-adapter, retraining

   ✓Quiz · 5 questions→