Model Online / On-Site
Duration
Modules 5 modules

Course Objective & Scope

Teams often green-light fine-tuning based solely on the cheap price of a single training run, overlooking the real costs: data curation, hosting a dedicated GPU endpoint 24/7, and re-training every time a new foundation model releases. This course teaches how to quantify quality, cost and risk before committing to custom models.

Participants model per-token API pricing against dedicated GPU serving (vLLM/SGLang or cloud provisioned throughput), evaluate catastrophic forgetting and safety degradation risks, and plan for model obsolescence. The deliverable is a TCO spreadsheet and risk mitigation plan for an internal initiative.

Who Should Attend?

  • AI engineers and technical leads preparing feasibility cases for custom LLM projects
  • ML platform and FinOps specialists managing cloud and GPU budgets
  • Engineering managers evaluating ROI and risk across enterprise AI initiatives
  • Prerequisite: familiarity with LLM deployment concepts and cloud cost fundamentals

Key Takeaways

  • Build a complete TCO model covering data curation, experimentation, evaluation and 24/7 inference serving
  • Calculate break-even traffic thresholds across serverless APIs, multi-LoRA serving and dedicated GPU instances
  • Weigh task-specific quality gains against general capability regressions and alignment tax
  • Score operational risks including safety degradation, training data extraction and model obsolescence
  • Present an executive-ready feasibility report with clear break-even metrics and kill criteria

Curriculum

Full-Lifecycle TCO of LLM Adaptation
  • Below the waterline: expert annotation, synthetic data generation and failed experiment runs
  • Comparing one-off training GPU spend against 24/7 inference hosting costs
  • Economics of serving multiple fine-tuned adapters on shared GPUs via multi-LoRA
Quality Gains versus Regression Analysis
  • Translating marginal accuracy gains on the target task into concrete business value
  • Quantifying catastrophic forgetting and alignment tax on adjacent capabilities
  • Identifying when a small accuracy lift does not justify ongoing maintenance overhead
Security, Privacy and Compliance Risks
  • Accidental erosion of safety guardrails during supervised fine-tuning
  • Training data memorisation and membership inference leakage risks
  • Downstream deployer obligations under the EU AI Act and KVKK
Operational Sustainability and Model Obsolescence
  • Re-training technical debt when new generations of foundation models ship
  • Vendor lock-in on proprietary fine-tuning APIs versus portable open-weight adapters
Workshop: TCO Calculator and Risk Decision Case
  • Modelling 12-month TCO and traffic curves for an internal AI use case
  • Preparing a Go/No-Go decision brief with a quality, cost and risk scorecard

Other Courses in This Category

  • Baselines and Success Metrics: Pre-Adaptation Benchmarking for LLMs

    Course on building golden test sets, establishing zero/few-shot baselines, defining deterministic and LLM-as-a-judge metrics and setting acceptance thresholds before starting model adaptation.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Fine-Tuning Fundamentals: Introduction to Customising LLMs

    Foundational hands-on course on how LLM fine-tuning works: dataset formatting, chat templates, running a first adapter training job with Hugging Face TRL and Unsloth, and avoiding overfitting.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Advanced Fine-Tuning: Continual Pre-Training, Long Context and Model Merging

    Advanced course covering continual pre-training (CPT), RoPE long-context extension, multi-turn tool-calling fine-tuning, catastrophic forgetting mitigation and model merging (SLERP, TIES, DARE) with MergeKit.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Supervised Fine-Tuning (SFT): Engineering and Optimisation

    Deep-dive engineering course on Supervised Fine-Tuning (SFT): completion-only loss masking, sequence packing, FlashAttention-2, NEFTune, learning rate schedules and production workflows with TRL SFTTrainer and Axolotl.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Parameter-Efficient Fine-Tuning (PEFT): Architectures and Multi-Adapter Serving

    Course on PEFT architectures—LoRA, DoRA, rsLoRA, VeRA, LoftQ, IA3 and Prefix Tuning—VRAM memory budgeting and serving dozens of dynamic adapters on a single base model with vLLM and SGLang.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View
  • Practical LLM Fine-Tuning with LoRA and QLoRA

    Hands-on training on Low-Rank Adaptation (LoRA) and 4-bit NormalFloat (NF4) QLoRA: rank/alpha tuning, target modules, double quantisation, paged optimizers and safe adapter merging.

    LLM Customisation & Model Adaptation Corporate
    Live Course
    5 modules
    View

Get in touch