What is a model parameter?
Begin with the mental picture, then use the technical details when they help.
Picture a large mixing board covered in small dials. Training shows the model example after example and nudges many dials at once. A parameter is one dial's learned numerical setting. Together, the settings shape how the model responds.
The one-sentence answer
A parameter is a numerical value inside a model that training is allowed to adjust, commonly stored in tensors such as weights and biases.
The mechanics
A small equation shows where learned values appear.
A tiny model
For y = wx + b, x is the current input, y is the output, and learned w and b are parameters.
y = wx + b
xinput
wweight
bbias
youtput
Training loop
- Receive a training example.
- Predict with the current parameters.
- Measure error with a loss function.
- Use backpropagation to calculate adjustment directions.
- Let an optimizer nudge many parameters, then repeat.
One example can affect many parameters, and one parameter can be affected by many examples. There is no normal one-line-to-one-parameter mapping.
Parameter versus hyperparameter
A parameter is learned. A hyperparameter is chosen, such as learning rate, batch size, training steps, or LoRA rank.
Where the information is
The bridge from Session 1 to Session 2's Data Traffic Light.
| Layer | What it is | Remember |
|---|---|---|
| Training data | Examples used in pretraining, fine-tuning, or adapter training. | Examples influence updates. Each example is not one parameter. |
| Parameters | Learned tensors such as weights, biases, and embedding tables. | They shape behavior collectively; one value rarely has readable meaning. |
| Prompt and context | Text, images, files, retrieval, or tool results supplied for this task. | Runtime input. Retention depends on the product, plan, and settings. |
| Current activations | Intermediate representations computed while processing the input. | Calculations for the run, not newly trained parameters. |
| Output, logs, history | Results and records saved by an app, connector, or organization. | Storage, access, deletion, and disclosure are governance questions. |
Privacy boundary: A prompt is not a parameter, but that does not make it safe to upload.
Check what is allowed, which tool is approved, what the exact settings retain, and who reviews the result.
Embeddings and text encoding
A learned embedding table contains parameters. The vector computed for one current prompt is an activation for that run. These ideas are related, but they are not interchangeable.
Memorization
Parameters are not a readable database of training lines, yet models can sometimes reproduce unusual or repeated sequences. Not stored as a row does not mean impossible to recover or automatically private.
How LoRA changes the picture
The same kind of learned numbers, with a smaller training target.
Full fine-tuning changes many base-model parameters. Low-Rank Adaptation usually freezes them and trains smaller low-rank matrices for selected weights.
Conceptual form
effective weight = base weight + scaled LoRA update
A LoRA contains learned adapter parameters and configuration metadata, not one training caption per parameter. Rank, targets, scale, data, and training choices matter.
More parameters can increase capacity, file size, memory use, and computation. Parameter count is not a direct intelligence score.
Common mix-ups
Four distinctions that prevent common errors.
- A parameter is a training line.An example helps cause many updates; the example and learned values differ.
- A parameter is a logic decision.It is a number. Behavior emerges from many numbers, operations, and the current input.
- An embedding is always a parameter.A learned table contains parameters; a current input embedding is computed for that run.
- No literal rows means no privacy risk.Memorization, retention, logs, integrations, and settings remain separate risks.
How the lab uses this
Session 1: separate a training example, learned setting, and current prompt.
Session 2: locate where organizational information enters or remains, then apply the Data Traffic Light.
Use the analogy in the presentation. Keep this page available for learners who want the technical detail.
Technical sources
Primary documentation and research papers.
- PyTorch: Parameter
- LoRA paper
- Hugging Face PEFT: LoRA
- Extracting Training Data from Large Language Models
Product retention, training, account, and privacy behavior can change. Check the exact service, plan, and settings before delivery.