Fine-Tuning and Model Adaptation

A decision framework based on whether you need knowledge, behaviour or format, with cost comparisons.

Assumes: Prompt Engineering Fundamentals

Updating every parameter: memory arithmetic for weights, gradients and optimiser states, and when it is justified.

Assumes: Prompt, RAG or Fine-Tune?

Why fine-tuning erases prior capability, how to measure it, and replay and regularisation mitigations.

Assumes: Full Fine-Tuning

Turning a base model into an assistant: task diversity, response quality and template consistency.

Assumes: Full Fine-Tuning

Completion vs chat formats, special tokens, loss masking on prompts, and template mismatch bugs.

Assumes: Instruction Tuning

Additive, selective and reparameterisation methods mapped, with trainable-parameter comparisons.

Assumes: Catastrophic Forgetting

The low-rank update derived, why it works, the zero-initialisation detail, and merging adapters at inference.

Assumes: Parameter-Efficient Fine-Tuning: The Landscape · Singular Value Decomposition

Choosing rank, alpha scaling, target modules, dropout and learning rate, with practical defaults.

Assumes: LoRA: Low-Rank Adaptation

Bottleneck adapters inserted in the block, the inference-latency cost, and adapter composition.

Assumes: Parameter-Efficient Fine-Tuning: The Landscape

Learning soft tokens instead of weights, prefix tuning of the KV cache, and their capacity limits.

Assumes: Adapter Layers

Floating point formats, scale and zero-point, symmetric vs asymmetric, per-tensor vs per-channel.

Assumes: Parameter-Efficient Fine-Tuning: The Landscape

Outlier features and LLM.int8(), 4-bit formats, NF4's information-theoretic motivation, and quality loss.

Assumes: Quantisation Fundamentals

Post-training quantisation algorithms, activation-aware weighting, and the GGUF deployment format.

Assumes: INT8, INT4 and NF4

4-bit NF4 quantisation, double quantisation, paged optimisers, and fine-tuning large models on one GPU.

Assumes: LoRA Hyperparameters · INT8, INT4 and NF4

Weight-decomposed adaptation, rank-stabilised scaling, LoRA+, VeRA and when variants actually help.

Assumes: QLoRA

FP16 vs BF16, loss scaling, master weights, and where numerical instability arises.

Assumes: Quantisation Fundamentals

Trading compute for memory, activation checkpointing arithmetic, and gradient accumulation.

Assumes: Mixed Precision Training

Data, tensor, pipeline and sequence parallelism, collective operations, and communication cost.

Assumes: Gradient Checkpointing and Memory Optimisation

Sharding parameters, gradients and optimiser states across stages 1 to 3, with memory arithmetic.

Assumes: Distributed Training

Quality over quantity, deduplication, decontamination, diversity measurement and annotation guidelines.

Assumes: Dataset Formats and Chat Templates

Self-Instruct and Evol-Instruct, distillation from stronger models, filtering, and model-collapse risk.

Assumes: Data Curation for Fine-Tuning

Building preference pairs, running DPO end to end, beta tuning, and diagnosing reward over-optimisation.

Assumes: Synthetic Data Generation · DPO and Direct Preference Optimisation

Teacher–student training, soft targets and temperature, and sequence-level distillation for LLMs.

Assumes: Synthetic Data Generation

Magnitude, structured and movement pruning, the lottery ticket hypothesis, and hardware-friendly sparsity.

Assumes: Knowledge Distillation

Task-specific evals, regression suites against the base model, held-out sets and human review.

Assumes: Preference Tuning in Practice

Overfitting small datasets, template mismatch, degenerate repetition, loss spikes and lost safety behaviour.

Assumes: Evaluating a Fine-Tuned Model