Session
Tailoring LLMs for specific domains
Thursday 23 July
11:00 – 12:00
Educator 2
Back to agenda
Generic LLMs are powerful but inefficient when applied to specialized tasks — they consume more resources than necessary and offer no competitive differentiation. This talk explores the model rearchitecting pipeline: a systematic approach to transforming generic models into optimized, domain-specific solutions. We'll cover three key phases:
- Structural Optimization (Pruning) — Surgically remove model components that contribute least to your target task, reducing size and increasing inference speed.
- Knowledge Recovery (Distillation) — The pruned "student" model learns to replicate the original "teacher" model's reasoning patterns, recovering lost capabilities efficiently.
- Specialization (Fine-tuning with LoRA) — Adapt the recovered model to a specific domain using parameter-efficient techniques that train only a small subset of weights.
For the demo, we will consider a specific use case to walk through the full pipeline end-to-end, demonstrating how each phase builds on the previous one. By the end, you'll understand how to move beyond prompt engineering and fine-tuning to architect smaller, faster, and more accurate models tailored to your needs.
- Structural Optimization (Pruning) — Surgically remove model components that contribute least to your target task, reducing size and increasing inference speed.
- Knowledge Recovery (Distillation) — The pruned "student" model learns to replicate the original "teacher" model's reasoning patterns, recovering lost capabilities efficiently.
- Specialization (Fine-tuning with LoRA) — Adapt the recovered model to a specific domain using parameter-efficient techniques that train only a small subset of weights.
For the demo, we will consider a specific use case to walk through the full pipeline end-to-end, demonstrating how each phase builds on the previous one. By the end, you'll understand how to move beyond prompt engineering and fine-tuning to architect smaller, faster, and more accurate models tailored to your needs.
