Generation Playground

What is this?

The Generation Playground lets you see exactly how the Fireworks API tokenizes your prompt before it reaches the model, and inspect the per-token logprobs of the completion. It's pre-loaded with a prompt injection demo that shows how safe_tokenization defends against control-token smuggling.

  1. Paste your Fireworks API key in the Configuration panel below.
  2. Hit ▶ Run — the playground sends the request with echo: true and logprobs auto-injected.
  3. Explore the results:
    • What the Model Sees — every prompt token as a chip. Amber chips are control tokens; gray chips are text. Hover for token IDs.
    • Generated Tokens — completion tokens colored by confidence (green = high, red = low). Hover for logprobs & top-5 alternatives.
  4. Toggle Safe Tokenization on/off and re-run to compare how the tokenization — and the model's answer — changes.
Configuration

Auto-injected: raw_output, logprobs, top_logprobs: 5, return_token_ids, echo, perf_metrics_in_response