Skip to contents

R package · on CRAN · MIT licensed

Fine-tune a language model
from R. In minutes.

Drop in a table, drag its columns into Prompt and Response, press Train. dragonfarm teaches a small model your format and your domain, on your own GPU or a free cloud one.

install.packages("dragonfarm")

Get started View on GitHub

dragon_app()

Step 1, Data: the example support tickets loaded into the appStep 2, Map: subject and body dragged into the Prompt slot, reply into ResponseStep 3, Model: a base model chosen and the GPU detectedStep 4, Train: training settings and the Start training buttonStep 5, Monitor: the loss curve of a finished runStep 6, Try it: the fine-tuned and base model replies side by side

Same model, same ticket

Teach it your house style.

A 0.5B-parameter model, before and after training on 190 example support tickets. This ticket was not in the training data. It is the median of the 10 held-out tickets by word overlap with the reference reply, and both replies are at temperature 0.

TicketLumen bulb pack arrived damaged. Opened my Lumen bulb pack today and the casing is split. The packaging looked fine from the outside so it must have happened before shipping.
Base modelI’m sorry to hear that your Lumen bulb pack has been damaged. It’s important to take any damage seriously and report it to the manufacturer or seller as soon as possible. Here are some steps you can follow: 1. Contact the Manufacturer: Reach out to the company that sold you the Lumen bulb pack. They may be able to provide a replacement or offer a repair service. 2. Report the Damage: … (shortened)
After fine-tuningI’m sorry your Lumen bulb pack arrived damaged. We will replace it at no cost. Please reply with a photo of the damage and one of the shipping label so we can file the carrier claim on our side. As soon as we have those, we ship a replacement Lumen bulb pack with a prepaid return label for the damaged unit. Usually a business day after the return is scanned in.Policy and tone are right. The last sentence mixes in a detail from another ticket type (the real policy is two business days), which is why small models still need review.

The example tickets are synthetic and built from templates, so this shows a model picking up a house style and policy, not open-ended reliability. Use your own data and check the output before relying on it.

Three ways in

Start where you’re comfortable.

Everything the app does is also an R function, and every run is a plain folder of files you can inspect.

Code

Five lines

Map your columns, train, generate. 

library(dragonfarm)
run <- dragon_dataset(dragon_example_data()) |>
  dragon_map(prompt = "{subject}\n\n{body}",
             response = "reply") |>
  dragon_train("Qwen/Qwen2.5-0.5B-Instruct",
               wait = TRUE)
dragon_generate(run,
  "My thermostat keeps dropping off Wi-Fi.")
App

Drag and drop

Six numbered steps, from loading a table to comparing against the base model. No code needed. 

library(dragonfarm)
dragon_app()
Cloud

No GPU? No problem.

Send the run to a free Colab or Kaggle GPU and bring the adapter home. Colab trains the example in about four minutes. 

run <- dragon_bundle(tickets,
  "Qwen/Qwen2.5-0.5B-Instruct")
dragon_remote(run, "colab")
dragon_import(run, "results.zip")

Go further

When one pass isn’t enough.

Preference tuning, reinforcement learning with rewards you can check, judges, chat, and publishing to the Hugging Face Hub are all in the box.

Fine-tune, judge, prefer, repeat

DPO and ORPO on top of a fine-tuned run, with a judge that checks the result.

Rewards you can check

GRPO with exact, numeric, JSON, length, and test-runner rewards.

Every function, documented

The full reference, from dragon_check() to dragon_publish().

Your first fine-tune is ten minutes away.

Install the package, run dragon_check(), and open the app. The first run sets up Python for you.

install.packages("dragonfarm")

Get started

CRANGitHubNewsReport an issue