R2 Stage D · 160 h (48 T / 112 P)
GenAI / LLM Engineer
Adapts, serves and evaluates language models; owns fine-tuning and inference performance.
Tech stack
PyTorchHugging Face TransformersDatasetsPEFTTRLbitsandbytesvLLMTGIDeepEvalWeights & BiasesMLflow
Modules
M1 · Dataset engineering
36 h TOPICS
- Task definition
- Instruction dataset construction
- Synthetic data generation and filtering
- Quality rubrics
- Deduplication and decontamination
- Train/validation/test hygiene
- Licensing and provenance
- Annotation operations
Lab: Construct, clean and document an instruction dataset with a held-out evaluation split
Course material for this module is in production.
M2 · Adaptation techniques
44 h TOPICS
- Prompt engineering versus RAG versus fine-tuning decision framework
- Supervised fine-tuning
- LoRA and QLoRA: rank, alpha, target modules
- Preference optimisation (DPO)
- Catastrophic forgetting
- Hyperparameter selection
- Compute and memory planning
Lab: Fine-tune an open-weights model with a parameter-efficient method and version the artefact
Course material for this module is in production.
M3 · Inference optimisation and serving
44 h TOPICS
- Quantisation formats and quality impact
- KV-cache mechanics and reuse
- Paged attention
- Continuous batching
- Speculative decoding
- Tensor parallelism
- GPU memory budgeting and utilisation
- Throughput versus latency tuning
- Autoscaling and cold start
Lab: Serve the tuned model; benchmark tokens/second, p95 latency, memory and cost per million tokens against the baseline
Course material for this module is in production.
M4 · Rigorous evaluation and release
36 h TOPICS
- Task-specific benchmark design
- Base-versus-tuned comparison
- Statistical significance
- Safety and regression evaluation
- Model cards
- Staged rollout and rollback criteria
Lab: Publish a comparative evaluation report supporting a release decision
Course material for this module is in production.
Track project
Domain-adapted language model: build the dataset, fine-tune with a parameter-efficient method, self-host on an optimised serving stack, and publish a benchmark report versus the base model covering quality, latency, memory and cost, with a documented release recommendation.
Job-ready exit standard
Can fine-tune, serve and evaluate models; explains quantisation, batching and KV-cache trade-offs with measured evidence.