TUTOR

M15 · Level III — AI Systems Engineering

Foundation Models & Fine-Tuning

Understand how broad pretrained models are formed and how they are adapted to specialised tasks, domains, modalities and system requirements.

Topics

  • Pretraining
  • Foundation Representation
  • Adaptation
  • Fine-Tuning
  • Instruction
  • Preference
  • Parameter Efficiency
  • Evaluation
  • Deployment

Technical concepts

  • 01The Foundation-Model Paradigm
  • 02Transfer Learning as a Statistical Assumption
  • 03Choosing the Adaptation Boundary
  • 04Domain Adaptation and Distribution Shift
  • 05Catastrophic Forgetting
  • 06Supervised Fine-Tuning
  • 07Instruction Tuning and Task Conditioning
  • 08Training-Data Curation as Model Engineering
  • 09Parameter-Efficient Fine-Tuning
  • 10Low-Rank Adaptation: Mathematical Derivation
  • 11Low-Rank Updates as Constrained Optimisation
  • 12Freezing Policies and Layer-Wise Adaptation
  • 13Fine-Tuning Learning Rates and Update Scale
  • 14Regularisation During Adaptation
  • 15Instruction Dataset Design
  • 16Preference Data and Relative Feedback
  • 17Preference Optimisation: Conceptual Frameworks
  • 18Alignment as an Objective-Design Problem
  • 19Evaluating Fine-Tuned Models
  • 20Benchmark Leakage and Validation Contamination
  • 21Multi-Task and Mixture Fine-Tuning
  • 22Continual and Sequential Adaptation Bridge
  • 23Quantisation-Aware Adaptation and Low-Bit Deployment Bridge
  • 24Multimodal Foundation Models and Adaptation
  • 25Foundation Models as Components of Agentic Systems
  • 26Security of Adapted Models
  • 27Quantum Technology Bridge: Foundation Models and Hybrid Adaptation

Practical implementation

  • 01Laboratory I: Baseline Transfer and Fine-Tuning Matrix
  • 02Laboratory II: Scratch Low-Rank Adaptation
  • 03Laboratory III: Instruction-Data Curation Engine
  • 04Laboratory IV: Supervised Fine-Tuning
  • 05Laboratory V: Instruction-Tuning Mixture Study
  • 06Laboratory VI: Catastrophic Forgetting and Retention
  • 07Laboratory VII: Preference-Data Adaptation
  • 08Laboratory VIII: Low-Bit Adaptation Study
  • 09Laboratory IX: Foundation-Model Evaluation Harness
  • 10Laboratory X: Multimodal Adaptation

Tools and technologies

  • Linux VM/container
  • parameter-efficient experiments
  • controlled fine-tuning
  • evaluation harnesses
  • reproducible checkpoints

Project

Adapted model + evaluation harness + release packet

Expected learning outcomes

  • 01Can define a foundation-model adaptation problem and boundary.
  • 02Can compare frozen, partial and full adaptation.
  • 03Can derive and implement a low-rank adaptation mechanism.
  • 04Can curate and audit instruction-tuning data.
  • 05Can quantify catastrophic forgetting and test a mitigation.
  • 06Can evaluate adapted models beyond the training objective.
  • 07Can audit benchmark contamination and reproducibility.
  • 08Can measure resource cost and checkpoint state.
  • 09Can defend adaptation choices and limitations in a technical viva.

Assessment

  • 01Daily practice
  • 02Laboratories
  • 03Research reproduction
  • 04Integrated project
  • 05Technical viva
  • 06Engineering record

Source document: TUTOR Frontier Module 15 Foundation Models Fine Tuning