Experiment Recipes

Recipes are self-contained YAML files under examples/, grouped by training domain. Each recipe is the source of truth for model, algorithm, rollout engine, placement, reward, synchronization, and batch geometry. Select one with --config-name=<path-under-examples> and omit .yaml.

Entrypoint per Domain

The defaults below come directly from each entrypoint's @hydra.main; they are useful compose-check anchors, not universal recommendations.

Training path	Entrypoint	Canonical recipe
Diffusion RL	`python -m unirl.train_diffusion`	`diffusion/sd3/sd3_trainside`
AR/VLM RL	`python -m unirl.train_ar`	`ar/qwen_vl_grpo_geo3k_mc_4x8`
SFT	`python -m unirl.train_sft`	`sft/qwen3_sft`
Async AR RL	`python -m unirl.train_async_ar`	`ar/qwen3_grpo_4b_base_dapo_sglang_async`
Async diffusion RL	`python -m unirl.train_async_diffusion`	`diffusion/bagel/bagel_vllmomni_async`
Prompt enhancement	`python -m unirl.train_pe`	`pe/pe_trainside_pickscore`
Unified AR + diffusion	`python -m unirl.train_unified_model`	`unified_model/hi3_vllmomni`
Agentic tool use	`python -m unirl.train_agentic`	`deep_research/deep_research_search_judge`

Choosing a Recipe Family

The tree currently contains more than 140 recipes, so this page intentionally does not duplicate a complete filename inventory. Inspect the matching directory and start from these stable families:

examples/diffusion/: image, video, audio-video, and image-editing RL. Model families such as SD3, Qwen-Image, WAN, HunyuanVideo, BAGEL, and LTX-2 usually have their own nested directory.
examples/ar/: Qwen text, VLM, Qwen3-Omni, and BAGEL autoregressive RL.
examples/sft/: text, VLM, agent trajectory, diffusion, Cosmos3, and WAN SFT.
examples/pe/: prompt-enhancer workflows.
examples/unified_model/: mixed AR + diffusion models.
examples/deep_research/: service-scored agentic workflows.

Selecting a Recipe

python -m unirl.train_diffusion \
  --config-name=diffusion/sd3/sd3_trainside \
  --cfg job --resolve

Launchers pass the same path. Set ENTRY for a non-diffusion entrypoint:

bash examples/run_experiment_single_node.sh diffusion/sd3/sd3_trainside
ENTRY=train_ar bash examples/run_experiment_single_node.sh ar/qwen_vl_grpo_geo3k_mc_4x8
ENTRY=train_sft bash examples/run_experiment_single_node.sh sft/qwen3_sft
ENTRY=train_agentic bash examples/run_experiment_single_node.sh deep_research/deep_research_search_judge
bash examples/run_experiment_multinode.sh diffusion/sd3/sd3_sglang_rollout_colocate

How to Pick a Recipe

Use this decision order when a task does not name a specific recipe:

Choose the training path and model family.
Choose direct trainside sampling or a dedicated SGLang/vLLM-Omni engine.
Match algorithm, adapter, synchronization, and placement suffixes.
Match topology-specific variants such as 4x8, VeOmni SP/EP, or HSDP.
Read the YAML and run a compose check before launching Ray.

Editing Guidance

When adding a recipe:

Start from the closest existing YAML under the correct domain/model directory.
Keep model, reward, rollout engine, backend, stack, sync, placement, and batch geometry in YAML, each instantiated by _target_.
Use environment interpolation only for deployment-specific paths and logging identity.
Compose it with the entrypoint that owns its training domain.