Skip to main content
In the previous guide, we cover doing RFT on a simple math example running locally. In this guide, we’ll show you how you can run RFT in your production environment with a remote rollout server, by walking through fine-tuning for an image (SVG) generation agent. Watch a quick walkthrough:

​
What You’ll Learn

  • Apply RFT to production agents — Train models that work with remote servers and existing infrastructure
  • Remote rollout processing — Connect your production environment to Fireworks RFT using Eval Protocol
  • Monitor and debug training — Track progress, inspect rollouts, and debug issues with live logs

​
1. Installation

  1. Clone the quickstart repo: https://github.com/eval-protocol/quickstart
git clone git@github.com:eval-protocol/quickstart.git
cd quickstart
  1. Install Eval Protocol:
pip install "eval-protocol[svgbench]"
  1. Environment Setup:
Make a copy of env.example, name it .env, and fill in the keys: 
FIREWORKS_API_KEY=your-fireworks-key-here
OPENAI_API_KEY=your-openai-key-here
Place this file in your evaluator directory (e.g., evaluator/.env). The create process below automatically reads and uploads these secrets to Fireworks.

​
2. Test your evaluator locally

Test your evaluator locally before launching training, to verify everything works with your rollout processor. Terminal 1 - Start the local UI server to view results: 
ep logs
Terminal 2 - Kick off the test: 
pytest evaluator/test_svgagent.py -vs
The test automatically uses our Vercel remote server: 
rollout_processor=RemoteRolloutProcessor(
    remote_base_url="https://vercel-svg-server-ts.vercel.app",
)
If you want to use a local development Vercel server instead, see Local Development Server.

​
Expected Test Output

The test should automatically open a browser page to view results. If it doesn’t, navigate to http://localhost:8000. 
INFO:eval_protocol.pytest.remote_rollout_processor:Found status log for rollout democratic-way-12: Rollout democratic-way-12 completed
INFO:eval_protocol.pytest.remote_rollout_processor:Found Fireworks log for rollout democratic-way-12 with status code 100.0
INFO:eval_protocol.adapters.fireworks_tracing:Successfully converted 1 traces to evaluation rows | 3/8 [00:19<00:22, 4.52s/rollout]
...
Runs (Parallel): 100%|████████████████████████████████████████████| 1/1 [00:31<00:00, 31.07s/run]
PASSED
Eval Protocol Logs Interface If you’re interested in understanding how Remote Rollout Processing works and how it communicates with the remote server, see How Remote Rollout Processing Works.

​
3. Start training with a single command

To kickoff training, simply do: 
cd evaluator
eval-protocol create rft \
  --base-model accounts/fireworks/models/qwen3-0p6b
This command:
  1. Uploads secrets — reads your .env and uploads API keys as Fireworks secrets
  2. Uploads evaluator — packages and uploads your evaluation code
  3. Waits for build — polls evaluator status until ACTIVE (timeout: 10 minutes)
  4. Creates dataset — uploads your svgbench_dataset.jsonl
  5. Launches RFT job — starts reinforcement fine-tuning with your evaluator

​
Configuration & Troubleshooting

Training Parameters: We use Eval Protocol’s default values for training parameters (batch size, epochs, learning rate, LoRA rank, accelerator count, etc.). For a complete list of available RFT flags you can customize, see Fireworks RFT Command Documentation. Changing Evaluators: If you’ve made changes to your evaluator code and want to upload a new version: 
eval-protocol create rft \
  --base-model accounts/fireworks/models/qwen3-0p6b \
  --force
Evaluator Upload Timing Out: If your evaluator takes longer than 10 minutes to build, you’ll see: 
⏰ Timeout after 10.0m - evaluator is not yet ACTIVE

❌ Evaluator is not ready within the timeout period.
📊 Please check the evaluator status at: https://app.fireworks.ai/dashboard/evaluators/test-svgagent-test-svg-generation-evaluation
   Wait for it to become ACTIVE, then run 'eval-protocol create rft' again.
In this case, monitor the evaluator upload at the link, and run the command again when ACTIVE.

​
4. Monitor Training Progress

After successful job creation, you’ll see: 
✅ Created Reinforcement Fine-tuning Job
   name: accounts/pyroworks/reinforcementFineTuningJobs/sdnld4yn

📊 Dashboard Links:
   Evaluator: https://app.fireworks.ai/dashboard/evaluators/test-svgagent-test-svg-generation-evaluation
   Dataset:   https://app.fireworks.ai/dashboard/datasets/svgbench-dataset
   RFT Job:   https://app.fireworks.ai/dashboard/fine-tuning/reinforcement/sdnld4yn
Click on the RFT Job link to view real-time training progress, epoch counts, and rollout data.

​
Training Results

After successful training, you should see performance improvements reflected in the training metrics: SVG Agent Training Progress

​
SVG Quality Improvement

You can inspect individual rollouts to see the dramatic improvement in SVG generation quality. Below is a comparison between the first epoch and the final 8th epoch: Before (1st Epoch): SVG Generation - Before Training After (8th Epoch): SVG Generation - After Training The reinforcement fine tuning process significantly improves the model’s ability to generate accurate, detailed SVG graphics that better match the input descriptions.

​
Debugging Tips

When your training is running, you have several powerful tools to debug and monitor your rollouts:

​
Rollout Overview

Clicking on any Epoch or Step in the training dashboard, then clicking the table icon to the right, will show you a comprehensive table of all rollouts. It’s a good high-level overview to see if any rollouts failed and for what reason. Rollout Overview Table

​
Individual Rollout Details

If you click on a specific row in the rollout table, you can see exactly what the prompt was and how the model responded. You can even copy and paste out the SVG code generated and render it yourself to see what the model did. This is how we got the results above in the before and after comparison. Individual Rollout Details

​
Live Log Streaming

Clicking on View Logs takes you to a page of logs being streamed in. Here, you can see precisely what errors are happening to the rollouts. This is useful to debug and fix any issues with your rollouts. Live Log Streaming

​
Contact Us / Learn More

​
Appendix

​
How Remote Rollout Processing Works

Eval Protocol enables reinforcement learning that meets you where you are. Instead of forcing you to rewrite your agent in a specific framework, you can implement a lightweight remote server wherever your codebase and infrastructure already live. Your remote server is only responsible for:
  • Executing rollouts - Run your agent logic (in this case, SVG generation from text prompts)
  • Logging to tracing - Send structured logs to tracing.fireworks.ai for evaluation (see the below linked docs for more information)
In this example, we showcase a Vercel TypeScript server that executes single-turn SVG code generation.
📖 Learn More: For a complete deep-dive into Remote Rollout Processing, see the Remote Rollout Processor Tutorial.

​
Local Development Server

cd vercel_svg_server_ts
vercel dev
Then swap out the remote_base_url to point to the local server you just started: 
rollout_processor=RemoteRolloutProcessor(
    remote_base_url="http://localhost:3000",
)
And in a third terminal, run the evaluation: 
pytest evaluator/test_svgagent.py -vs
See Vercel CLI documentation for more information on local development.