Skip to content
IconMind

AI & LLM

Training icons

61 icons in the training group of AI & LLM: Activation, Adapter, Add checkpoint and 58 more — each in outline and duotone at three weights, with the code for React, Vue, Svelte, Flutter and more.

61 icons · 1 of 10 groupsOpen in browser

The 61 training icons in AI & LLM, described

Activationactivation
An activation — nothing, nothing, then everything, the nonlinear fire of a ReLU.
Adapteradapter
An adapter — a small trained piece like LoRA bolted on to a big frozen model.
Add checkpointadd-checkpoint
A flag flown big, a plus on its banner — save a new checkpoint right now.
Alignmentalignment
Alignment — bringing a model round to what people actually want, steering with feedback.
Calibrationcalibration
Points scattered close to the identity line — how well a model's confidence matches its accuracy.
Checkpointcheckpoint
A checkpoint — a model's weights saved exactly as they were at this moment in training.
Checkpoint alertcheckpoint-alert
A flag flown big, an alert on its banner — a checkpoint that needs attention.
Checkpoint comparecheckpoint-compare
Checkpoint compare — this saved state against that one, to pick the better.
Checkpoint loadcheckpoint-load
Checkpoint load — pick up exactly where training stopped, restored from a save.
Checkpoint savecheckpoint-save
Checkpoint save — write the training state down before going on.
Datasetdataset
A dataset — the collection of examples, rows or documents a model is trained and evaluated on.
Distillationdistil
Distillation — a smaller student model taught to imitate a bigger teacher.
DPOdpo
DPO — direct preference optimisation, training a model on which of two answers people preferred.
Dropoutdropout
Dropout — some neurons sit this batch out at random to regularise the model.
Early stoppingearly-stop
Early stopping — halt training before the model starts getting worse.
Epochepoch
An epoch — one full pass over all the training data during a training run.
Eval harnesseval-harness
A play button and a check inside one frame — the harness that runs evaluations and grades them.
Find checkpointfind-checkpoint
A flag flown big, a lens on its banner — find the checkpoint you need.
Fine-tunefine-tune
Fine-tune — nudge a model with a small extra pass rather than retraining it.
Fine-tuningfine-tuning
Fine-tuning — adapting an already trained model to one domain, task or style with extra data.
Frozen checkpointfrozen-checkpoint
A flag flown big, a padlock on its banner — a checkpoint frozen so it cannot change.
Frozen layerfrozen-layer
A frozen layer — this layer keeps its weights fixed during transfer learning.
Gradient flowgrad-flow
Gradient flow — the correction travelling back down through the layers in backprop.
Gradientgradient
A gradient — the slope that says which way is downhill for the optimiser.
Gradient clipgradient-clip
Gradient clipping — no single update step gets to be bigger than the ceiling.
Gradient descentgradient-descent
Gradient descent — walk downhill on the loss until there is no downhill left.
Grokkinggrokking
A curve that stays flat and then suddenly climbs — the moment a model finally generalises.
Labelled checkpointlabelled-checkpoint
A flag flown big, a label on its banner — a checkpoint named so it can be found.
Learning ratelearning-rate
The learning rate — how big a step the optimiser takes, often decayed on a schedule.
LoRA mergelora-merge
LoRA merge — the adapter folded back into the base weights and baked in.
Lossloss
Loss — how wrong the model still is, the objective that training drives down.
Loss curveloss-curve
A loss curve — the training error coming down over time as a model converges.
LR schedulelr-schedule
An LR schedule — the learning rate stepped down on a plan as training proceeds.
Model checkpointmodel-checkpoint
A model core beside a flag — a checkpoint saved during training.
Model goalmodel-goal
A model core beside a target — the objective a model is trained toward.
Model trendmodel-trend
A model core beside a rising line — a model's quality trending over versions.
Next milestonenext-milestone
A flag flown big, a double chevron on its banner — on to the next milestone.
Noisenoise
Noise — random scatter and jitter that carries no signal worth keeping.
Noise schedulenoise-schedule
Dots shrinking step by step along a diagonal — the noise schedule of a diffusion process.
Normalisenormalize
Normalise — bring values onto the same scale so they can be compared.
Optimiseroptimiser
An optimiser — the algorithm like Adam that walks the loss downhill to a minimum.
Overfitoverfit
Overfit — a curve that followed every wobble of the training data and memorised it.
Overfit gapoverfit-gap
The overfit gap — training loss keeps falling while validation turns away.
Precision-recallprecision-recall
A curve that holds high then falls away — the precision-recall trade-off of a classifier.
Pre-trainingpretrain
Pre-training — the first and biggest pass over everything, building the base model.
Pruneprune
Prune — cut away the parameters or branches that were not earning their place.
Quantizationquantization
Quantization — using fewer bits per weight on purpose so a model runs smaller and faster.
Rank adapterrank-adapter
A rank adapter — the small LoRA piece bolted on to a big frozen model.
Regulariseregularise
Regularise — hold a model back from memorising with dropout or a penalty.
Remove checkpointremove-checkpoint
A flag flown big, a minus on its banner — delete a checkpoint you no longer need.
Resume checkpointresume-checkpoint
A flag flown big, a play on its banner — resume training from a checkpoint.
RLHFrlhf
RLHF — reinforcement learning from human feedback, a person inside the training loop.
ROC curveroc-curve
A curve bowing above the chance diagonal — the ROC curve of a classifier.
Stale checkpointstale-checkpoint
A flag flown big, a Z on its banner — a checkpoint too old to be useful any more.
Patiencestop-patience
Patience — the plateau is on the clock before early stopping ends the run.
Training losstrain-loss
Training loss — down, then flat, the curve every run hopes to draw.
Trainingtraining
Training — teaching a model from data by adjusting its weights over many epochs.
Training datatraining-data
Training data — the labelled examples and corpus a model learned from.
Underfitunderfit
Underfit — a line too simple for the shape it is chasing, high bias that misses the pattern.
Underfitunderfit-gap
Underfit — both curves stay high, the model has not learned enough yet.
Warmupwarmup
Warmup — the learning rate ramps gently up to speed, then holds.

In code, each is one import — import { Activation } from "@iconmind/react/icons/activation" — and the same name in Vue, Svelte, Solid, Preact, React Native, Astro, Blade and Flutter.

Narrow it by tag

Every training icon in AI & LLM, free to ship

61 icons in outline and duotone at three weights, generated from one grid so nothing in the set can drift out of step. MIT licensed — commercial use, no attribution, no seat count.

Other groups in AI & LLM