Man pages for dragonfarm
Fine-Tune Small Language Models with LoRA from R
| dragon_app | Launch the dragon-farm app |
| dragon_archived_runs | Archived runs |
| dragon_archive_run | Archive, restore, or delete a run |
| dragon_backend | Inference backends |
| dragon_bundle | Package a run for a cloud GPU |
| dragon_cancel | Cancel a run |
| dragon_chat | Talk to a model, with memory of the conversation |
| dragon_check | Check the Python environment and hardware |
| dragon_code | R code that reproduces a run |
| dragon_compare | Compare runs side by side |
| dragon_conversations | Multi-turn conversations as training data |
| dragon_dataset | Create a dataset for fine-tuning |
| dragon_evaluate | Evaluate a finished run |
| dragon_example_data | Path to the bundled example dataset |
| dragon_export_gguf | Export a merged model to GGUF |
| dragon_feedback | Training data from chat feedback |
| dragon_generate | Generate replies from a fine-tuned model |
| dragon_hardware | Hardware settings |
| dragon_import | Import results trained on another machine |
| dragon_judge | Judge a run's replies with a language model |
| dragon_llm_anthropic | Language models as functions: the Claude API and ellmer |
| dragon_lora | LoRA settings |
| dragon_map | Map dataset columns to prompt, response, and system text |
| dragon_map_pairs | Map columns for preference optimization |
| dragon_map_prompts | Map columns for reinforcement learning |
| dragon_merge | Merge the adapter into the base model |
| dragon_metrics | Task metrics for generated replies |
| dragon_pipeline | Run several post-training stages as one pipeline |
| dragon_pipeline_cancel | Cancel a pipeline |
| dragon_pipeline_status | Progress of a pipeline |
| dragon_prefer | Preference optimization with DPO or ORPO |
| dragon_presets | Recommended small models |
| dragon_preview | Preview mapped rows as chat turns |
| dragon_prompts | Prompts from a run's data files |
| dragon_publish | Push a run's model to the Hugging Face Hub |
| dragon_python_requirements | Python requirements used by dragonfarm |
| dragon_reinforce | Reinforcement learning with verifiable rewards (GRPO) |
| dragon_remote | Open a cloud GPU provider for a bundled run |
| dragon_remote_providers | Cloud GPU providers |
| dragon_resume | Resume a run from its latest checkpoint |
| dragon_reward | Verifiable rewards for reinforcement learning |
| dragon_run | Reopen an existing run |
| dragon_runs | List runs |
| dragon_runs_dir | Directory where runs are stored |
| dragon_serve_ollama | Serve a run's model with Ollama |
| dragon_split | Hold out rows for evaluation |
| dragon_status | Inspect a run |
| dragon_step | Steps of a post-training pipeline |
| dragon_synthesize | Write fine-tuning data with a teacher model |
| dragon_synthesize_pairs | Build preference pairs from a model's own samples |
| dragon_train | Fine-tune a model with LoRA |
| dragon_train_args | Training settings |
| dragon_wait | Wait for a run to finish |
| dragon_worker_stop | Stop the local inference worker |
