Package: dragonfarm 0.3.4

dragonfarm: Fine-Tune Small Language Models with LoRA from R

Fine-tune small (100M to 3B parameter) causal language models with LoRA (Low-Rank Adaptation) from R. Datasets are mapped to chat-format prompts and responses, training runs in a background 'Python' process built on Hugging Face 'transformers' and 'peft', and a 'shiny' app offers drag-and-drop dataset upload and column mapping. 'Python' dependencies are declared through 'reticulate' and resolved automatically on first use.

Authors:Tejas Patel [aut, cre]

dragonfarm_0.3.4.tar.gz
dragonfarm_0.3.4.zip(r-4.7-any)dragonfarm_0.3.4.zip(r-4.6-any)dragonfarm_0.3.4.zip(r-4.5-any)
dragonfarm_0.3.4.tgz(r-4.6-any)dragonfarm_0.3.4.tgz(r-4.5-any)
dragonfarm_0.3.4.tar.gz(r-4.7-any)dragonfarm_0.3.4.tar.gz(r-4.6-any)
dragonfarm_0.3.4.tgz(r-4.6-emscripten)
manual.pdf |manual.html✨
DESCRIPTION |NEWS
card.svg |card.png
dragonfarm/json (API)

# Install 'dragonfarm' in R:
install.packages('dragonfarm', repos = c('https://algocrat.r-universe.dev', 'https://cloud.r-project.org'))

Bug tracker:https://github.com/tejas4patel/dragon-farm/issues

On CRAN:

Conda:

3.60 score 518 downloads 73 exports 87 dependencies

Last updated from:40b84c89b9. Checks:9 OK. Indexed: yes.

TargetResultTimeFilesSyslog
linux-develOK178
source / vignettesOK363
linux-releaseOK176
macos-releaseOK229
macos-oldrelOK215
windows-develOK145
windows-releaseOK140
windows-oldrelOK149
wasm-releaseOK117

Exports:dragon_appdragon_archive_rundragon_archived_runsdragon_backenddragon_backend_localdragon_backend_ollamadragon_backend_serverdragon_bundledragon_canceldragon_chatdragon_chat_loaddragon_checkdragon_codedragon_comparedragon_conversationsdragon_datasetdragon_delete_rundragon_evaluatedragon_example_datadragon_export_ggufdragon_feedbackdragon_generatedragon_hardwaredragon_importdragon_judgedragon_judge_anthropicdragon_judge_ellmerdragon_llm_anthropicdragon_llm_ellmerdragon_logsdragon_loradragon_mapdragon_map_pairsdragon_map_promptsdragon_mergedragon_metric_regexdragon_metricsdragon_pipelinedragon_pipeline_canceldragon_pipeline_statusdragon_preferdragon_presetsdragon_previewdragon_progressdragon_promptsdragon_publishdragon_python_requirementsdragon_reinforcedragon_remotedragon_remote_providersdragon_resumedragon_rewarddragon_rundragon_runsdragon_runs_dirdragon_serve_ollamadragon_splitdragon_statusdragon_step_evaluatedragon_step_judgedragon_step_mergedragon_step_preferdragon_step_publishdragon_step_reinforcedragon_step_synthesize_pairsdragon_step_traindragon_synthesizedragon_synthesize_pairsdragon_traindragon_train_argsdragon_unarchive_rundragon_waitdragon_worker_stop

Dependencies:askpassassertthatbackportsbase64encbslibcachemcheckmateclicommonmarkcpp11crosstalkcurldata.tabledigestdplyrevaluatefarverfastmapfontawesomefsgenericsggplot2gluegtableherehighrhtmltoolshtmlwidgetshttpuvhttrisobandjquerylibjsonliteknitrlabelinglaterlatticelazyevallearnrlifecyclelitedownmagrittrmarkdownMatrixmemoisemimeopensslotelpillarpkgconfigplotlypngprocessxpromisespspurrrR6rappdirsRColorBrewerRcppRcppTOMLrenvreticulaterlangrmarkdownrprojrootS7sassscalesshinysortablesourcetoolsstringistringrsystibbletidyrtidyselecttinytexutf8vctrsviridisLitewithrxfunxtableyamlzip

Quickstart: fine-tune a small model from R
1. Check the machine | 2. Load and map the data | 3. Train | 4. Evaluate and try it | 5. Ship it | Choosing settings

Last update: 2026-10-07
Started: 2026-10-07

Reinforcement learning with rewards you can check
When it helps, and when it does not | Data: prompts and references | Rewards | Train | Read the result | Settings that matter | Chaining with the other stages

Last update: 2026-10-07
Started: 2026-10-07

The dragon-farm app
The eight panels | Running it for other people | Under the hood

Last update: 2026-10-07
Started: 2026-10-07

The post-training loop: fine-tune, judge, prefer, repeat
The shape of the loop | 1. Fine-tune | 2. Measure | 3. Make preference pairs from the model's own replies | 4. Preference-optimize on top of the fine-tune | 5. Compare, then go around again | The same loop in one call | Talk to the result | When this loop is the wrong tool

Last update: 2026-10-07
Started: 2026-10-07

Readme and manuals

Help Manual

Help pageTopics
Launch the dragon-farm appdragon_app
Archive, restore, or delete a rundragon_archive_run dragon_delete_run dragon_unarchive_run
Archived runsdragon_archived_runs
Inference backendsdragon_backend dragon_backend_local dragon_backend_ollama dragon_backend_server
Package a run for a cloud GPUdragon_bundle
Cancel a rundragon_cancel
Talk to a model, with memory of the conversationdragon_chat dragon_chat_load
Check the Python environment and hardwaredragon_check
R code that reproduces a rundragon_code
Compare runs side by sidedragon_compare
Multi-turn conversations as training datadragon_conversations
Create a dataset for fine-tuningdragon_dataset dragon_dataset.character dragon_dataset.data.frame
Evaluate a finished rundragon_evaluate
Path to the bundled example datasetdragon_example_data
Export a merged model to GGUFdragon_export_gguf
Training data from chat feedbackdragon_feedback
Generate replies from a fine-tuned modeldragon_generate
Hardware settingsdragon_hardware
Import results trained on another machinedragon_import
Judge a run's replies with a language modeldragon_judge
Language models as functions: the Claude API and ellmerdragon_judge_anthropic dragon_judge_ellmer dragon_llm_anthropic dragon_llm_ellmer
LoRA settingsdragon_lora
Map dataset columns to prompt, response, and system textdragon_map
Map columns for preference optimizationdragon_map_pairs
Map columns for reinforcement learningdragon_map_prompts
Merge the adapter into the base modeldragon_merge
Task metrics for generated repliesdragon_metrics dragon_metric_regex
Run several post-training stages as one pipelinedragon_pipeline
Cancel a pipelinedragon_pipeline_cancel
Progress of a pipelinedragon_pipeline_status
Preference optimization with DPO or ORPOdragon_prefer
Recommended small modelsdragon_presets
Preview mapped rows as chat turnsdragon_preview
Prompts from a run's data filesdragon_prompts
Push a run's model to the Hugging Face Hubdragon_publish
Python requirements used by dragonfarmdragon_python_requirements
Reinforcement learning with verifiable rewards (GRPO)dragon_reinforce
Open a cloud GPU provider for a bundled rundragon_remote
Cloud GPU providersdragon_remote_providers
Resume a run from its latest checkpointdragon_resume
Verifiable rewards for reinforcement learningdragon_reward
Reopen an existing rundragon_run
List runsdragon_runs
Directory where runs are storeddragon_runs_dir
Serve a run's model with Ollamadragon_serve_ollama
Hold out rows for evaluationdragon_split
Inspect a rundragon_logs dragon_progress dragon_status
Steps of a post-training pipelinedragon_step dragon_step_evaluate dragon_step_judge dragon_step_merge dragon_step_prefer dragon_step_publish dragon_step_reinforce dragon_step_synthesize_pairs dragon_step_train
Write fine-tuning data with a teacher modeldragon_synthesize
Build preference pairs from a model's own samplesdragon_synthesize_pairs
Fine-tune a model with LoRAdragon_train
Training settingsdragon_train_args
Wait for a run to finishdragon_wait
Stop the local inference workerdragon_worker_stop