README.md
In dragonfarm: Fine-Tune Small Language Models with LoRA from R

dragonfarm

Fine-tune small language models (roughly 100M to 3B parameters) with LoRA, from a CLI or a few lines of Python. This is the Python side of dragon-farm, which also ships an R package with a drag-and-drop Shiny app; the two share this same code, so a run started from R can be inspected or continued from Python and back.

pip install dragonfarm

# Environment and hardware, as JSON
dragonfarm check

# Run one post-training stage (sft, preference optimization, or GRPO,
# chosen by the run directory's config.json) on a run someone already set up
dragonfarm train --run-dir path/to/run

# Generate replies from a request file: {"model": ..., "adapter": ...,
# "prompts": [...], ...} in, {"outputs": [...]} out
dragonfarm generate --request request.json --out reply.json

# Zip a finished run's adapter, config, and samples for import elsewhere
dragonfarm pack path/to/run

Run dragonfarm <command> --help for each command's full set of flags.

A run directory (the same one the CLI and the R package both read and write) can be inspected without touching torch or transformers at all:

from dragonfarm.api import Run, runs

run = Run("dragonfarm_runs/20260101-120000-my-model")
run.state          # "running", "succeeded", "failed", ...
run.stage          # "sft", "prefer", or "reinforce"
run.config         # the full config.json as a dict
run.progress()     # one row per logged training step
run.log(tail=20)   # the last 20 lines of the trainer's output

for r in runs("dragonfarm_runs"):
    print(r.id, r.stage, r.state)

Every run is a plain directory of JSON and JSONL files: config.json (model, LoRA, and stage settings), data/train.jsonl and data/eval.jsonl, status.json, progress.jsonl, log.txt, and, once training finishes, adapter/ (the LoRA weights) and optionally merged/ (the adapter folded into the base model). Nothing here is specific to R or to Python; either can create, read, or continue a run.

MIT