RL Export

What exists today

The page is reachable only by typing its URL: /{org}/{project}/rl-export. It is titled RL Export, “Label trajectories or export evaluator-scored traces as JSONL for fine-tuning and RLHF”, and has two tabs.

Tab	What you see	What actually happens
Label	One fixed example transcript and a Rate this trajectory panel with 1 to 5 sliders for Correctness, Helpfulness, Efficiency and Safety, a computed Reward (the four ratings summed and divided by 20), and Skip / Save & next.	The transcript is built-in sample text, not one of your traces. Skip and Save & next do nothing. Ratings are lost when you leave.
Export	Toggles for reward source (Human labels / Evaluator score) and format (Trajectory / Prompt/completion), stat cards (Records, Avg reward, Pass rate, Human-labeled), a Preview, and Download .jsonl.	The toggles don’t change anything. The stats are fixed numbers. Download .jsonl saves the same single sample record shown in the preview.

The sample record shows the shape the feature is meant to produce, one JSON object per trajectory:

{
  "reward": 0.5,
  "passed": false,
  "rubric": { "correctness": 3, "helpfulness": 3, "efficiency": 3, "safety": 3 },
  "messages": [
    { "role": "user", "content": "Plan my wellness goals" },
    { "role": "assistant", "content": "Here is a balanced plan..." }
  ],
  "agent_version": "v4",
  "scenario": "happy_path_wellness_and_social_goals"
}

Although the preview is labelled .jsonl, the downloaded sample is a single pretty-printed JSON object spread over several lines, so it is not valid JSON Lines.

Getting your data out today

Until RL Export works, these are the real ways to get scored data out of AgentHub:

You want	Use
A readable write-up of a run, evaluation or trace	Markdown or PDF export on any report.
Per-evaluator scores, pass/fail and judge reasoning for a trace	The evaluate API’s `GET` on the trace, which returns every saved trace evaluation run with its results.
The raw spans of a trace	The Raw JSON section of each span on the trace detail page, or the traces API.
A fabricated dataset for testing	The CSV download per table in Synthetic Setup.

Get started

Platform

Evaluate and observe

Administration

What exists today

Getting your data out today

​What exists today

​Getting your data out today

What exists today

Getting your data out today