Learn
Practical guides to fine-tuning, distillation, and deploying small language models.
What Are Tokens and Tokenization?
Tokenization is the subword split that turns text into the integers a model consumes. It decides how many tokens your particular text costs, and it varies by model.
What Is a Context Window?
The context window is the hard token budget shared by prompt, history and output. Exceeding it is an error, and filling it isn't free even when you stay inside.
What Is a KV Cache?
The KV cache stores attention keys and values from earlier tokens so generation stays linear per step, and it's why serving memory climbs as a conversation runs on.
Catastrophic Forgetting: Short Definition
A one-page definition of catastrophic forgetting: where the name comes from, how it differs from overfitting, and where to find the diagnosis-and-fix version.
LoRA Rank: Short Definition
A one-page definition of LoRA rank, covering which config key it is, what moving it does, and where the arithmetic and tuning advice live.
What Is Model Quantization?
Quantization stores model weights at lower numeric precision to cut memory, a deliberate accuracy-for-footprint trade you make after training rather than instead of it.
What Is Temperature in LLMs?
Temperature divides the logits before softmax, setting how sharply a model commits to its top token, and on distil labs its valid range depends on which teacher family you picked.