Overfitting
Memorising the training data instead of learning the pattern.
When a model fits its training examples so closely that it performs badly on anything new. The training loss looks great and real-world results do not.
The same failure appears in prompt and eval work: a prompt tuned against ten examples that crumples on the eleventh. Hold out data you never look at.
In LLM work it shows up in three places: fine-tunes that memorise your examples, prompts tuned until they only work on the cases you looked at, and retrieval tuned against a handful of queries. All three share the same signature — excellent scores on the material you developed against, and a surprise in production.
The prompt tuned into a corner
A prompt scores perfectly on the twelve examples used to refine it.
Hold back a set of cases you never look at until the end. The gap between the tuned set and the held-out set is the honest measure of whether the work generalised — and it is usually wider than expected.
- guideIntroduction to large language models ↗
A plain-language official introduction to models, tokens and context. Teaching material rather than origin research — good for the definition, not for a claim about the frontier.
- paperLoRA: low-rank adaptation of large language models ↗
Freeze the pretrained weights, train small low-rank matrices beside them. The origin of parameter-efficient fine-tuning, and why adapters are cheap to ship and swap.
Evaluation · Evals
How you know the system is actually working.
Golden Dataset · Eval set
A curated set of cases with known-good answers.
Regression Test
A test that catches a specific past failure coming back.
Catastrophic Forgetting
Fine-tuning that improves one thing and breaks others.