# Playground

### Arkane Cloud playground

Arkane Cloud provides a *model playground* where users can test various AI models [available in Arkane Cloud Playground](https://console.arkanecloud.com/playground) through a web interface without coding.

### How to use the playground

1. On the [Models](https://console.arkanecloud.com/models) page, choose a model and click **Deploy** on its card.
2. In **System prompt**, enter instructions that the model should follow when generating output. You can include examples of good in-context learning outputs (few-shot prompting). For example:

   > Add a fun fact about trains to the end of each response. The fun fact may not be related to what I'm asking you about. Here is an example:
   >
   > \- User: What is the capital of Croatia?\
   > \- You: The capital of Croatia is Zagreb. By the way, did you know that the first railroad in New England was powered by horses?
3. Start chatting to the model.

### Model parameters

The playground allows you to set basic sampling parameters for the model you have chosen.

For more parameters supported by the [inference API](https://docs.arkanecloud.com/arkane-cloud/api-reference/api-reference), refer to the [vLLM documentation](https://docs.vllm.ai/en/stable/dev/sampling_params.html#vllm.SamplingParams).

#### Temperature

* **Affects**: Output randomness
* **Models**: All
* **Type**: Number
* **Range of values**: From 0 to 1 for meta-llama/Meta-Llama-3, from 0 to 2 for models in other families
* **Default value**: From 0.3 to 0.7, depending on the model

Temperature determines how "hot-headed" the model predictions are. The higher the temperature, the more random and less deterministic and conservative the output.

> For example, a 0.8 temperature makes outputs more creative and random than a 0.5.

#### Max tokens

* **Affects**: Computational cost
* **Models**: All
* **Type**: Number
* **Range of values**: Depends on the model (Usually between 1 to 4096 tokens)
* **Default value**: Depends on the model

The maximum number of tokens that the model generates. 

#### Presence penalty

* **Affects**: Output randomness
* **Models**: All
* **Type**: Number
* **Range of values**: From −2 to 2
* **Default value**: 0

The presence penalty is applied to new tokens that have previously appeared in the output. Positive values penalize such tokens and negative values favor them.

#### Top-p threshold

* **Affects**: Output randomness
* **Models**: All
* **Type**: Number
* **Range of values**: From 0 to 1
* **Default value**: From 0.9 to 1, depending on the model

In *top-p sampling*, also known as *nucleus sampling*, the model considers only the most probable tokens whose combined probability mass is equal to the specified threshold.

> For example, with a threshold of 0.1, only the tokens that comprise the top 10% of the probability mass are considered.

To consider all tokens, set the threshold to 1.