5.2
5.2 Supervised Fine-Tuning & Adaptation
LoRA/QLoRA/DoRA low-rank adaptation, loss masking, multi-task data mixtures, curriculum learning, and offline quality regression gates.
LoRA, QLoRA, DoRA & High-Throughput Fine-Tuning
LoRA low-rank decomposition math ($W = W_0 + \frac{\alpha}{r}BA$), QLoRA NormalFloat4 (NF4) quantized fine-tuning, DoRA magnitude-direction decoupling, loss masking, and Unsloth/PEFT fast kernels.
🏢 Companies
Unsloth AI、Hugging Face (PEFT)、Predibase、Together AI、Databricks
🛠️ Tech Stack
LoRAQLoRADoRASFTLoss MaskingPEFTUnsloth
💼 Roles & Salary
Fine-Tuning Engineer、ML Engineer、LLM Application Engineer
💰 $195K - $390K / year (SFT & PEFT Engineering) | ¥500K - ¥1.2M / year
📚 Prerequisites: Matrix Rank & SVD Linear Algebra • PyTorch Autograd & Backpropagation • LoRA Hyperparameter (r, alpha) Tuning • Hugging Face PEFT & Trainer Internals
Instruction Mixture, Curriculum Tuning & Regression Gates
Multi-domain instruction weighting, difficulty-graded curriculum fine-tuning, catastrophic forgetting mitigation, chat template format safety, and automated pre-deployment regression gates.
🏢 Companies
Hugging Face、Scale AI、Databricks、Snorkel AI
🛠️ Tech Stack
Instruction MixtureCurriculum TuningCatastrophic ForgettingChat TemplateRegression GatesEvaluation Gate
💼 Roles & Salary
Fine-Tuning Engineer、Evaluation Engineer、ML Engineer
💰 $185K - $370K / year (Instruction & Eval Gates) | ¥480K - ¥1.1M / year
📚 Prerequisites: Instruction Mixture & Ablation Experiments • Catastrophic Forgetting Mitigation (Replay) • Chat Template Tokenization & Safety • Automated Benchmark Regression Suites