Skip to main content
The Playground lets you talk to any of your endpoints before you wire it into a product. It is the quickest way to check how a fine-tuned version or a base model answers.
Valkyrie Playground screen

The Playground: a conversation with an endpoint, with a system prompt set on the right.

Chat with an endpoint

1

Pick an endpoint

Choose one of your running endpoints in the Endpoint card. Base-model endpoints are marked (base). Open endpoint takes you to its page.
2

Adjust the settings (optional)

Add a System prompt with instructions for every reply, and set the Temperature and Max tokens.
3

Send a message

Type a message and press Enter to send it. Shift+Enter adds a new line. Click Clear to start over.
If you have no running endpoint, the Playground shows links to deploy one. See Endpoints.
Getting an empty reply? Some models think before they answer, and that thinking counts toward Max tokens. If the limit is low, the model can use it all up before it writes the answer. Raise Max tokens (the default of 512 is enough for most questions) and send the message again.

Compare side by side

Click Compare side by side to send every message to two endpoints at once, with the same settings, and read the answers in two columns. Each answer shows how long the model took, for example Answered in 1.8 s, so you can compare speed as well as quality. What you compare against, under Compare with:
  • This endpoint, without fine-tuning. The default for a fine-tuned version served as a LoRA adapter on vLLM. The same endpoint answers with the adapter switched off, so you see what fine-tuning changed without deploying the base model on a second GPU.
  • Another endpoint. For a base model, or a fine-tune served as merged weights, Valkyrie picks a running endpoint of the same base model for you. You can choose any other endpoint from the list. If none is running, deploy the base model first.
Messages you send here are real requests to your endpoints, so they must be running. In compare mode both answers are billed like any request, including the one without fine-tuning. The adapter-off answer is a Playground feature: it isn’t available for fine-tunes served as merged weights or through the OpenAI-compatible API.