Skip to main content
This feature is in BETA. The API and behavior may change in future releases.

​
Overview

Reinforcement Fine-Tuning (RFT) allows you to fine-tune language models using reinforcement learning on HUD tasks. This creates specialized models optimized for your specific use cases.
Access Required: RFT is currently available by invite only. Contact founders@hud.ai to request access.
RFT currently supports OpenAI models and requires a minimum of 10 tasks for effective training.

​
Prerequisites

  • RFT access (contact founders@hud.ai to request)
  • HUD API key set in environment (HUD_API_KEY)
  • A HUD environment with remote tasks configured
  • At least 10 tasks in your dataset
  • No vision support required (RFT doesn’t support vision-based environments)

​
Step-by-Step Process

​
1. Prepare Your Tasks

First, ensure you have a tasks file with at least 10 tasks. You can download existing datasets or create your own: 
# Download a dataset
hud get hud-evals/SheetBench-50

# Or use your existing tasks
ls tasks.json

​
2. Launch RFT Training

Run the RFT command from within your environment directory: 
hud rft run tasks.json
The command will submit the job to the RFT service
Use --yes flag to skip all confirmation prompts for automation.

​
3. Monitor Training Progress

Check the status of your training job: 
# Get status using the model ID from the launch output
hud rft status <model_id>

# View detailed status information
hud rft status <model_id> --verbose

​
4. Use Your Fine-Tuned Model

Once training is complete (status shows “succeeded”), you’ll receive a fine-tuned model ID. Use it with the HUD inference API: 
import os
from openai import OpenAI

client = OpenAI(
    base_url="https://inference.hud.ai", 
    api_key=os.getenv("HUD_API_KEY")
)

response = client.chat.completions.create(
    model="openai/ft:o4-mini-2025-04-16:human-union-data::CdqkVdJy",
    messages=[
        {"role": "user", "content": "Say hello!"}
    ],
)

print(response.choices[0].message.content)
Replace the model ID in the example with your actual fine-tuned model ID from the status command.

​
Command Reference

​
Launch Training

hud rft run <tasks_file> [OPTIONS]

Options:
  --provider TEXT           Provider to use (default: openai)
  --reasoning-effort TEXT   Reasoning effort level: low, medium, high (default: medium)
  --verbose / -v           Enable verbose output
  --yes / -y               Auto-accept all prompts

​
Check Status

hud rft status <model_id> [OPTIONS]

Options:
  --verbose / -v   Show full status details

​
Tips

  • Start with smaller datasets (10-50 tasks) to test your approach
  • Use --verbose during development to see detailed information
  • Monitor logs for any errors during training
  • Fine-tuned models are optimized for tasks similar to your training data

​
Limitations

  • Minimum 10 tasks required
  • No vision support (text-based tasks only)
  • Currently supports OpenAI models only
  • Training time varies based on dataset size

​
Troubleshooting

If training fails:
  1. Verify your tasks have valid remote configurations
  2. Ensure all environment variables are set
  3. Check that your environment doesn’t require vision support
  4. Use --verbose flag for detailed error messages