ELUVIXSovereign AI Router

Trace Playground & Fine-Tuning Datasets

The Trace Playground (available at /playground in the Next.js dashboard) is an interactive developer workbench for experimenting with prompts, testing routing pipelines, and curating fine-tuning datasets.

Interactive Chat & Trace Inspector

Unlike standard LLM chat interfaces, the Trace Playground exposes the complete internal execution metadata for every interaction:

  • Upstream Selected: Shows whether the query resolved to a local GPU or escalated to a cloud fallback.
  • Complexity Score: Displays the complexity grade (e.g. Tier 1: Low, Tier 3: Frontier) calculated by the complexity estimator.
  • Latency Breakdown: Measures end-to-end response time alongside TTFT (Time to First Token).
  • Guardrail Actions: Indicates whether the Privacy Guard redacted any PII tokens or whether the Semantic Guard performed safety validations.

Session Persistence

All playground conversations are automatically committed to the local SQLite database (chat_history table).

You can:

  • Switch between active conversation threads.
  • Inspect raw request and response payloads.
  • Re-run specific prompts with alternative temperature or system instruction parameters.

Exporting Fine-Tuning Datasets (JSONL)

High-quality local conversations can be converted into ready-to-train datasets with a single click or API call.

One-Click UI Export

In the Playground, click "Export Dataset" to download an .jsonl file formatted to the OpenAI Fine-Tuning standard:

{"messages": [{"role": "system", "content": "You are a helpful assistant."}, {"role": "user", "content": "Explain latency spillover."}, {"role": "assistant", "content": "Latency spillover routes requests..."}]}

Programmatic Export Endpoint

You can also fetch exported datasets via the REST API:

curl -k https://127.0.0.1:3000/api/v1/export \
  -H "Authorization: Bearer <ADMIN_PASSWORD>" \
  -o fine_tuning_dataset.jsonl

The resulting file can be fed directly into fine-tuning pipelines (such as Axolotl, Unsloth, LLaMA-Factory, or OpenAI fine-tuning) to distill cloud model outputs into small local models.