Reference

Glossary

What is this?

124 terms used in fine-tuning and serving language models, defined without marketing. Each entry leads with a definition that stands on its own, then explains what actually goes wrong with the thing in practice.

Training

Adapter

An adapter is the small set of trained weights produced by a parameter-efficient fine-tune.

Adapter merging

Adapter merging folds a trained adapter’s weights into the base model to produce a single standalone model.

Batch size

Batch size is the number of training examples processed before the model weights are updated.

Catastrophic forgetting

Catastrophic forgetting is the loss of previously learned general capability when a model is trained heavily on a narrow task.

Checkpoint

A checkpoint is a saved snapshot of training state — weights, optimiser state and position in the dataset — written periodically so a run can resume from it rather than restarting.

Distillation

Distillation trains a smaller model to reproduce the behaviour of a larger one.

Early stopping

Early stopping ends a training run when validation performance stops improving, rather than running to the planned number of steps.

Epoch

An epoch is one complete pass through the training dataset.

Fine-tuning

Fine-tuning is continuing to train an already-trained language model on a smaller, task-specific dataset so that it adapts to a particular domain, format or behaviour.

Full fine-tuning

Full fine-tuning updates every weight in the model rather than a small added set.

Gradient

A gradient is the derivative of the loss with respect to each trainable parameter — the direction and magnitude in which changing that parameter would most reduce the error.

Gradient accumulation

Gradient accumulation simulates a large batch on limited memory by processing several small batches, summing their gradients, and applying a single weight update at the end.

Gradient checkpointing · Activation checkpointing

Gradient checkpointing reduces memory by discarding intermediate activations during the forward pass and recomputing them during the backward pass.

Learning rate

The learning rate controls how large a step the optimiser takes on each update.

Learning rate schedule

A learning rate schedule varies the learning rate over the course of a run rather than holding it constant — usually a short warmup to full rate followed by a gradual decay.

LoRA · Low-Rank Adaptation

LoRA, or Low-Rank Adaptation, is a fine-tuning method that freezes the original model weights and trains small pairs of low-rank matrices injected alongside them.

Loss

Loss is a single number measuring how wrong the model’s predictions are on a batch.

Optimiser state

Optimiser state is the additional data an optimiser maintains per trainable parameter — typically running estimates of gradient mean and variance.

Overfitting

Overfitting is when a model memorises its training data rather than learning the pattern behind it.

Preference tuning · DPO, Direct Preference Optimisation

Preference tuning trains a model on pairs of responses where one has been marked better than the other, rather than on a single correct answer.

Prompt engineering

Prompt engineering shapes model behaviour through instructions supplied at request time.

QLoRA · Quantised LoRA

QLoRA is LoRA applied on top of a base model quantised to four bits.

Rank · LoRA rank, r

Rank is the inner dimension of the low-rank matrices an adapter trains, and it sets how much capacity the adapter has.

Reference model

A reference model is a frozen copy of the starting model, held during preference tuning to measure how far the trained model has drifted.

Reinforcement learning from verifiable rewards · GRPO, RLVR

A training approach where the model generates several candidate answers per prompt and is rewarded according to a programmatic check of their correctness.

Reward function

A reward function scores a model’s output during reinforcement learning.

Reward hacking

Reward hacking is a model finding a way to score highly under a reward function without doing the thing the function was meant to encourage.

Reward model

A reward model is a model trained to predict human preference, used to score outputs when no programmatic check exists.

Supervised fine-tuning · SFT

Supervised fine-tuning trains a model on pairs of inputs and desired outputs, teaching it to reproduce the target response for a given prompt.

Target modules

Target modules are the specific layers of a model that an adapter is attached to.

Training step

A training step is one weight update: a forward pass over a batch, a backward pass to compute gradients, and an application of those gradients by the optimiser.

Warmup

Warmup is a short period at the start of training during which the learning rate ramps from near zero to its full value.

Data

Alpaca format · Instruct format

The Alpaca format stores each example as an instruction, an optional input, and an output.

Chat template

A chat template is the model-specific rule for turning a list of role-tagged messages into the exact token sequence the model expects.

ChatML

ChatML is a conversation format that marks each message with an explicit role — system, user or assistant — using special tokens the model was trained to recognise.

Dataset validation

Dataset validation checks a training file before any compute is leased: that it parses, that every record has the required fields, that sequences fit the context window, and that duplicates and empty targets are flagged.

Deduplication

Deduplication removes repeated or near-repeated examples from a training set.

Format adapter

A format adapter converts a dataset from the shape you have into the shape the trainer requires — for example conversation records into a model’s native chat format, or preference records into training pairs.

Preference pair

A preference pair is a prompt with two responses, one marked better than the other.

Sequence length · Max sequence length

Sequence length is the maximum number of tokens in a single training example.

ShareGPT format

ShareGPT is a JSON conversation format storing a list of turns, each with a sender and a value.

Special tokens

Special tokens are reserved symbols that carry structural meaning rather than content — start and end of sequence, role boundaries, padding.

Synthetic data

Synthetic data is training data generated by a model rather than collected from real activity.

Token

A token is the atomic unit a language model reads and writes — typically a common word, a word fragment, or a punctuation mark.

Tokenisation

Tokenisation splits text into the discrete units a model actually processes.

Training example

A training example is one input-and-output pair the model learns from.

Truncation

Truncation is the silent cutting of an example that exceeds the configured sequence length.

Evaluation

Baseline model

The baseline is whatever is running today — the model a candidate would replace.

Benchmark

A benchmark is a standardised public evaluation set used to compare models.

Blinded evaluation

A blinded evaluation hides which model produced which output from whoever or whatever is judging.

Confidence interval

A confidence interval expresses the uncertainty around a measured result given the sample size.

Critical regression

A critical regression is a failure on a case the model must never get wrong — a safety refusal, a compliance requirement, a known past incident.

Data contamination

Data contamination is the presence of evaluation data in the training set, whether directly or through near-duplicates.

Evaluation gate

An evaluation gate is a mandatory check between training a model and deploying it: the candidate is scored against the model it would replace, and only a passing result permits the deploy.

Held-out set · Test set

A held-out set is data deliberately excluded from training and reserved for final measurement.

LLM-as-judge

LLM-as-judge uses a language model to grade the outputs of another model against a rubric.

Perplexity

Perplexity is the exponential of cross-entropy loss, interpretable as how many equally-likely options the model was effectively choosing between at each token.

Regression

A regression is a case where a new model performs worse than the one it replaces.

Rubric

A rubric is the explicit set of criteria a judge applies when scoring an output — what counts as correct, complete, appropriately formatted and appropriately toned.

Schema validity

Schema validity is the proportion of model outputs that parse and conform to the expected structure.

Statistical significance

Statistical significance is the judgement that an observed difference is unlikely to be an artefact of sampling.

Validation split

A validation split is a portion of data held out from training and evaluated periodically during a run, used to detect overfitting and to choose between checkpoints.

Win rate

Win rate is the proportion of evaluation cases where the candidate’s output was judged better than the baseline’s.

Serving

Agent

An agent is a system where a model plans and executes a sequence of tool calls toward a goal rather than answering in one turn.

Always warm

An always-warm endpoint keeps at least one replica resident at all times, so no request ever pays a cold start.

Batching

Batching processes multiple inference requests together in one pass.

Cold start

A cold start is the delay on the first request to an endpoint with no resident replica: capacity must be acquired, the model weights loaded and the runtime initialised.

Context window

The context window is the maximum number of tokens a model can attend to at once, covering the prompt and the generated output together.

Cooldown

Cooldown is how long an endpoint stays resident after its last request before releasing capacity.

Durable operation

A durable operation is a long-running job that survives client disconnection and process restarts, with an append-only event log a client can replay from any position.

Endpoint

An endpoint is a network address that serves a specific model, with its own scaling configuration and its own cost.

Idempotency

An idempotent operation produces the same result whether performed once or repeated.

Inference

Inference is running a trained model to produce output.

KV cache

The KV cache stores the key and value tensors computed for tokens already processed, so generating each new token does not require reprocessing the whole sequence.

Latency

Latency is the time between sending a request and receiving a response.

Multi-adapter serving

Multi-adapter serving hosts several fine-tuned adapters against a single resident copy of a base model, switching between them per request.

OpenAI-compatible API

An OpenAI-compatible API implements the same request and response shape as OpenAI’s chat completions endpoint.

Quantisation

Quantisation stores model weights at lower numerical precision — commonly eight or four bits instead of sixteen — to reduce memory and increase speed.

Rate limit

A rate limit caps how many requests a caller may make in a period.

Replica

A replica is one running copy of a model serving requests.

Resumption

Resumption is continuing an interrupted job from its last checkpoint rather than restarting it.

Retrieval-augmented generation · RAG

Retrieval-augmented generation fetches relevant documents at request time and places them in the model’s context.

Scale to zero

Scale to zero means an endpoint releases all its compute when idle and costs nothing until the next request.

Server-sent events · SSE

Server-sent events is a one-way streaming protocol over plain HTTP, where a server pushes a sequence of events to a client on a long-lived connection.

Streaming

Streaming returns output tokens as they are generated rather than waiting for the complete response.

Structured output

Structured output constrains a model to produce responses conforming to a schema, typically JSON.

Temperature

Temperature controls randomness in token selection.

Throughput

Throughput is how much work a deployment completes per unit time — tokens per second, or requests per second.

Time to first token · TTFT

Time to first token is the delay between a request arriving and the first output token being produced.

Tool calling · Function calling

Tool calling is a model producing a structured request to invoke an external function rather than answering in prose.

Cost and billing

Cost authorisation

Cost authorisation estimates the maximum cost of a job and reserves it against a spend limit before any compute is leased.

Entitlement

An entitlement is what an account is permitted to do under its current plan and subscription state — whether it may lease compute, up to what limit, with what concurrency.

GPU class

A GPU class is a family of interchangeable accelerators with the same memory capacity and comparable throughput, priced as a unit.

GPU-hour

A GPU-hour is one accelerator held for one hour, and it is the standard unit for compute pricing.

Metered billing

Metered billing charges for measured consumption rather than a fixed subscription.

Quota

A quota is a ceiling on a countable resource — concurrent jobs, concurrent endpoints, stored bytes.

Reservation

A reservation is budget held against a spend limit for a job that has been authorised but not yet completed.

Serverless compute

Serverless compute means capacity is acquired when work exists and released when it does not, with no cluster to provision or maintain.

Spend limit

A spend limit is a ceiling on what an account may consume over a period.

Usage event

A usage event is an immutable record of consumption — what was used, when, by whom, and under which price revision.

VRAM · GPU memory

VRAM is the memory on a GPU, and it is the binding constraint on almost every fine-tuning job.

Model architecture

Active parameters

Active parameters are the subset of a mixture-of-experts model actually used to process a given token.

Attention

Attention is the mechanism that lets a model weigh the relevance of every earlier token when processing the current one.

Base model

A base model is the pretrained model a fine-tune starts from.

Dense model

A dense model uses every parameter for every token.

Feed-forward network · MLP

The feed-forward network is the part of each transformer block that processes each position independently after attention has mixed information between positions.

Instruction-tuned model

An instruction-tuned model has already been fine-tuned to follow instructions and hold a conversation.

Mixture of experts · MoE

A mixture-of-experts model contains many specialised sub-networks and routes each token through only a few of them.

Model publishing

Model publishing pushes a trained model to a public or private hub so it can be discovered, downloaded and used elsewhere.

Model weights

Model weights are the learned parameters that constitute a trained model.

Open-weight model

An open-weight model has downloadable parameters, so it can be run, fine-tuned and deployed by anyone under its licence.

Parameter count

Parameter count is the number of learned weights in a model, usually quoted in billions.

Reasoning model

A reasoning model has been trained to produce explicit intermediate steps before its final answer.

Transformer

The transformer is the architecture behind essentially every current language model: stacked blocks of attention and feed-forward layers.

Put the vocabulary to work

A magic link creates your account and your first API key. Upload a dataset and see a validation report before spending anything.

Get your API key Read the docs