Skip to main content
🚧 Coming Soon - This page is under construction

​
Prerequisites

Placeholder: Before launching an RFT job, ensure you have:
  • Dataset uploaded to Fireworks
  • Evaluator uploaded via Eval Protocol
  • Fireworks API key configured
  • Sufficient GPU quota

​
Option A: CLI with Eval Protocol

Placeholder: Complete guide to launching via command line

​
Step 1: Install Eval Protocol CLI

Placeholder: Installation instructions 
pip install eval-protocol

​
Step 2: Authenticate

Placeholder: How to set up API credentials

​
Step 3: Upload Evaluator

Placeholder: Command to upload your evaluator 
eval-protocol upload --entry "module::function"
Example: Placeholder: Concrete example with real module name

​
Step 4: Create RFT Job

Placeholder: Full command with all options explained 
eval-protocol create rft \
  --base-model accounts/fireworks/models/llama-v3p1-8b-instruct \
  --dataset-id DATASET_ID \
  --evaluator-id EVALUATOR_ID \
  --output-model my-finetuned-model \
  --epochs 1 \
  --learning-rate 1e-4 \
  --inference-temperature 0.7 \
  --inference-n 4

​
Step 5: Verify Job Created

Placeholder: How to check job status 
eval-protocol list rft

​
Option B: Web UI

Placeholder: Complete guide to launching via dashboard

​
Step 1: Navigate to Fine-Tuning

Placeholder:
  1. Go to Fireworks Dashboard
  2. Click “Fine-Tuning” in sidebar
  3. Click “Fine-tune a Model”
Screenshot: Placeholder: Image of fine-tuning page

​
Step 2: Select Reinforcement Method

Placeholder:
  • Choose “Reinforcement” as tuning method
  • Select base model from dropdown
Screenshot: Placeholder: Image showing method selection

​
Step 3: Configure Dataset

Placeholder:
  • Upload new dataset or select existing
  • Preview dataset entries
  • Verify format
Screenshot: Placeholder: Image of dataset selection

​
Step 4: Select Evaluator

Placeholder:
  • Choose from uploaded evaluators
  • Preview evaluator code
  • Test on sample data
Screenshot: Placeholder: Image of evaluator selection

​
Step 5: Set Training Parameters

Placeholder: Form showing all parameters with descriptions:
  • Base model
  • Output model name
  • Epochs
  • Learning rate
  • LoRA rank
  • Max context length
  • Batch size
Screenshot: Placeholder: Image of parameter form

​
Step 6: Configure Rollout Parameters

Placeholder: Form for inference settings:
  • Temperature
  • Top-p
  • Top-k
  • Number of rollouts (n)
  • Max tokens
Screenshot: Placeholder: Image of rollout settings

​
Step 7: Review and Launch

Placeholder:
  • Review all settings
  • Estimated cost/time
  • Click “Start Fine-Tuning”
Screenshot: Placeholder: Image of review page

​
Using firectl CLI (Alternative)

Placeholder: For users who prefer firectl over eval-protocol 
firectl create rftj \
  --base-model llama-v3p1-8b-instruct \
  --dataset my-dataset \
  --evaluator my-evaluator \
  --output-model my-model

​
Comparing CLI vs UI

Placeholder: Table showing:
FeatureCLI (eval-protocol)UIfirectl
SpeedFastSlowerFast
AutomationEasyManualEasy
Parameter discoveryHarderEasierMedium
ReproducibilityExcellentManual trackingExcellent

​
Advanced Configuration

Placeholder: Less common options:
  • Custom GPU requirements
  • Environment URLs for multi-turn
  • Checkpoint frequency
  • W&B integration

​
Job Validation

Placeholder: How Fireworks validates your job before starting:
  • Dataset format check
  • Evaluator syntax check
  • Resource availability
  • Quota limits

​
Common Errors and Fixes

Placeholder: Error messages you might see:
  • “Invalid dataset format” → Fix dataset JSONL
  • “Evaluator not found” → Re-upload evaluator
  • “Insufficient quota” → Request more GPUs
  • “Invalid parameter range” → Check parameter bounds

​
After Launching

Placeholder: What happens next:
  1. Job enters queue
  2. Resources allocated
  3. Training starts
  4. You can monitor progress

​
Next Steps

Placeholder: Link to monitoring guide to track your job