Configuration

Config
type	object
properties
custom	Custom
	Custom script configuration.
	type	object
	additionalProperties	True
train	#/$defs/TrainingConfig
rollout	#/$defs/RolloutConfig
policy	#/$defs/PolicyConfig
logging	#/$defs/LoggingConfig
profiler	#/$defs/ProfilerConfig
validation	#/$defs/ValidationConfig
distillation	#/$defs/DistillationConfig
vla	#/$defs/VLAConfig
mode	Mode
	Running mode, could be ‘disaggregated’ or ‘colocated’
	type	string
	default	disaggregated
$defs
CheckpointConfig	CheckpointConfig
	type	object
	properties
	enable_checkpoint	Enable Checkpoint
		Enable checkpointing for training. If set to False, no checkpoint will be saved.
		type	boolean
		default	False
	save_freq	Save Freq
		Checkpoint save frequency for training steps
		type	integer
		default	20
	save_freq_in_epoch	Save Freq In Epoch
		Checkpoint save frequency for training epochs. Default to 0 (disabled).
		type	integer
		default	0
	save_mode	Save Mode
		Checkpoint save mode for training steps
		type	string
		default	async
	max_keep	Max Keep
		Maximum number of checkpoints to keep. If set to -1, all checkpoints will be kept.
		type	integer
		default	5
	export_safetensors	Export Safetensors
		Whether to export a safetensors weight for huggingface usage, include related config files.
		type	boolean
		default	True
	upload_hf	Upload Hf
		Whether to upload the safetensors weight to huggingface.
		type	boolean
		default	False
	hf_repo_name	Hf Repo Name
		The huggingface repo name to upload the safetensors weight.
		type	string
		default	Comos-Reason1
	upload_s3	Upload S3
		Whether to upload the checkpoint and safetensors to S3. Default to False, set final will upload the final checkpoint, all will upload all checkpoints.
		default	False
		anyOf	type	boolean
			type	string
	s3_bucket	S3 Bucket
		The S3 bucket name to upload the checkpoint and safetensors weight.
		default	None
		anyOf	type	string
			type	null
	s3_prefix	S3 Prefix
		The S3 prefix to upload the checkpoint and safetensors weight.
		type	string
		default	outputs
DatasetConfig	DatasetConfig
	type	object
	properties
	name	Name
		Huggingface dataset name or local path to parquet file
		type	string
		default
	subset	Subset
		Dataset subset if exists
		default
		anyOf	type	string
			type	null
	revision	Revision
		OrderedDict({‘help’: ‘Dataset git revision if exist, can be a branch name, a tag, or a commit hash.’})
		default
		anyOf	type	string
			type	null
	split	Split
		A list of dataset splits to train
		default
		anyOf	type	string
			type	array
			items	type	string
	test_size	Test Size
		Size of the test set. If float, it is the ratio (between 0.0 and 1.0) of the dataset; if int, it is the absolute size of the test set.
		default	None
		anyOf	type	number
			type	integer
			type	null
	local_dir	Local Dir
		Local path to load dataset
		type	string
		default
DiffusersConfig	DiffusersConfig
	type	object
	properties
	dtype	Dtype
		Data type for diffusers model
		type	string
		default	float16
	is_video	Is Video
		True if this model is video generate model
		type	boolean
		default	False
	max_prompt_length	Max Prompt Length
		Maximum sequence length to use for the prompt
		type	integer
		default	300
	weighting_scheme	Weighting Scheme
		Method used to sample timestep
		type	string
		default	logit_normal
	train_flow_shift	Train Flow Shift
		flow shift used for training
		type	number
		default	3.0
	offload	Offload
		Whether to dynamic offload model parts from cuda to cpu
		type	boolean
		default	True
	logit_mean	Logit Mean
		random sampling timestep logits mean for noise addition
		type	number
		default	0.0
	logit_std	Logit Std
		random sampling timestep logits std for noise addition
		type	number
		default	1.0
	inference_size	Inference Size
		Image/video size for generation, [height, width]
		type	array
		default	1024
			1024
		items	type	integer
	inference_frames	Inference Frames
		Total frame of video size for generation
		type	integer
		default	41
	train_frames	Train Frames
		Total frame of video size for training
		type	integer
		default	41
	timesteps_fraction	Timesteps Fraction
		Fraction of timesteps to use during training. if set to less than 1.0, the model will be trained on a subset of the timesteps for each sample. this will speed up training but reduce the accuracy of policy gradient estimates.
		type	number
		default	1.0
	weight_copy_decay_type	Weight Copy Decay Type
		Weight copy decay type for diffusers model in rl training
		type	integer
		default	0
	lora	LoRA configuration for diffusers model
		default	None
		anyOf	#/$defs/LoraConfig
			type	null
	sample	Sampling configuration
		#/$defs/SampleConfig
	tokenizer	#/$defs/TokenizerConfig
DistillationConfig	DistillationConfig
	type	object
	properties
	enable	Enable
		Whether to enable distillation.
		type	boolean
		default	False
	parallelism	#/$defs/ParallelismConfig
	model_name_or_path	Model Name Or Path
		The teacher model name or path, compatible with huggingface model name or local path
		type	string
		default	Qwen/Qwen2.5-VL-7B-Instruct
	model_revision	Model Revision
		The revision of the teacher model to use
		default	None
		anyOf	type	string
			type	null
	compile	Compile
		Whether to use torch.compile for teacher model.
		type	boolean
		default	True
	master_dtype	Master Dtype
		The master weight data type for teacher model, is orthognal to param_dtype. Should be high precision for convergence consideration
		type	string
		default	float32
	param_dtype	Param Dtype
		The data type for forward/backward of teacher model. Outside forward/backward, params are in master_dtype
		type	string
		default	bfloat16
	logprob_dtype	Logprob Dtype
		The data type for logprobs calculation of teacher model.
		type	string
		default	float32
	fsdp_reduce_dtype	Fsdp Reduce Dtype
		The data type for reduction in FSDP for teacher model.
		type	string
		default	float32
	fsdp_offload	Fsdp Offload
		Whether to offload the teacher model to CPU if using FSDP
		type	boolean
		default	False
	fsdp_reshard_after_forward	Fsdp Reshard After Forward
		Reshard the param after forward pass in FSDP for teacher model. Default to ‘never’ to avoid unnecessary overhead.
		type	string
		default	never
	batch_size_per_replica	Batch Size Per Replica
		Batch size for teacher model per replica.
		type	integer
		default	1
	max_token_len_per_mini_batch	Max Token Len Per Mini Batch
		Maximum token length per mini batch. If set, dynamic mini-batch sizing will be applied based on this limit for teacher model.
		default	None
		anyOf	type	integer
			type	null
	sequence_packing	Sequence Packing
		Whether to enable sequence packing for teacher model. If set to True, the input sequences will be packed into a single tensor for training stability.
		type	boolean
		default	False
	mini_batch	Mini Batch
		mini batch size for teacher model in each replica.
		type	integer
		default	2
	seed	Seed
		Random seed for teacher model.
default		None
anyOf		type	integer
		type	null
kl_penalty_coef	Kl Penalty Coef
	The coefficient for KL penalty.
	type	number
	default	1.0
kl_discount_factor	Kl Discount Factor
	The discount factor for KL penalty.
	type	number
	default	0.0
include_prompt	Include Prompt
	Whether to include prompt in the teacher model KL calculation.
	type	boolean
	default	False
top_k	Top K
	Top-k filtering for teacher model logits before KL calculation. If larger than 0, generalized Jensen-Shannon Divergence will be used.
	type	integer
	default	0
jsd_beta	Jsd Beta
	Interpolation coefficient between 0.0 and 1.0 of the Generalized Jensen-Shannon Divergence loss. When beta is 0.0, the loss is the KL divergence. When beta is 1.0, the loss is the Inverse KL Divergence.
	type	number
	default	0.5
trainer_token_ids_from_teacher	Trainer Token Ids From Teacher
	Whether the trainer gets all top_k token ids directly from its redis interacted teacher model during distillation rather than from the rollout structure. This can simplify the rollout payload when being transferred in the framework.
	type	boolean
	default	True
rollout_top_k_recompute	Rollout Top K Recompute
	Whether to recompute all top-k logprobs with top-k token ids after the full sequence generated during rollout for distillation. This can ensure the completion generation process with no large top-k kept so that not degrade the generation efficiency.
	type	boolean
	default	False
FP4Config	FP4Config
	type	object
	properties
	enable_fp4	Enable Fp4
		Whether to enable fp4.
		type	boolean
		default	False
	fp4_recipe	Fp4 Recipe
		Recipe for weight scale calculation.
		type	string
		default	dynamic_scaling
	quant_recipe	Quant Recipe
		Quantization strategy for weight.
		type	string
		default	rowwise
FP8Config	FP8Config
	type	object
	properties
	enable_fp8	Enable Fp8
		Whether to enable fp8.
		type	boolean
		default	False
	fp8_recipe	Fp8 Recipe
		Recipe for weight scale calculation.
		type	string
		default	dynamic_scaling
	quant_recipe	Quant Recipe
		Quantization strategy for weight.
		type	string
		default	rowwise
GrpoConfig	GrpoConfig
	type	object
	properties
	type	Type
		type	string
		const	grpo
	trainer_type	Trainer Type
		Type of the trainer for GRPO.
		type	string
		default	grpo
	variant	Variant
		Variant of the GRPO, currently support grpo, gspo, dapo
		type	string
		default	grpo
	dataset	Dataset configuration for GRPO training. It includes dataset name, subset, revision, train split, test split and test size.
		#/$defs/DatasetConfig
	dataloader_shuffle	Dataloader Shuffle
		Shuffle the dataloader. If False, the dataloader will be used in the order it is loaded.
		type	boolean
		default	True
	dataloader_seed	Dataloader Seed
		random seed for dataloader shuffling
		type	integer
		default	0
	data_dispatch_as_rank_in_mesh	Data Dispatch As Rank In Mesh
		Whether to dispatch data according to rank in global mesh. If True, each rank will get its specific data shard based on its rank in the global mesh.
		type	boolean
		default	False
	enable_dataset_cache	Enable Dataset Cache
		Enable dataset cache process results, maybe accelerate the dataset loading
		type	boolean
		default	False
	dataloader_num_workers	Dataloader Num Workers
		Number of subprocess to use for data loading
		type	integer
		default	0
	dataloader_prefetch_factor	Dataloader Prefetch Factor
		Number of batches loaded in advance by each worker.
		default	None
		anyOf	type	integer
			type	null
	dataloader_batch_size	Dataloader Batch Size
		Batch size for each iteration of the dataloader for when fetch prompts from controller. This is only the setting of the dataloader iterator on the controller side.
		default	1
		anyOf	type	integer
			type	null
	prompt_column_name	Prompt Column Name
		Column name for prompt
		type	string
		default
	response_column_name	Response Column Name
		Column name for response/reference answer
		type	string
		default
	reward_function	Reward Function
		Reward functions for the model. Currently support single_choice, boxed_math, and format. You can add weight to each reward function by passing a dict, e.g., {‘single_choice’: 0.9, ‘format’: 0.1}
		anyOf	type	string
			type	array
			items	type	string
			type	object
			additionalProperties	type	number
	use_remote_reward	Use Remote Reward
		Whether to use remote reward calculation. If set to True, the reward calculation will be done in a remote worker. If False, the reward calculation will be done in the local process.
		type	boolean
default		False
remote_reward	Configuration for remote reward calculation.
	#/$defs/RemoteRewardConfig
filter_reward_metric	Filter Reward Metric
	Reward function to filter in dynamic sampling for DAPO. If specified, only samples with different this rewards will be used for training. If None, no filtering will be applied.
	anyOf	type	string
		type	array
		items	type	string
bypass_reward	Bypass Reward
	Bypass reward computation and use fixed reward of 0.0 for all samples. Useful for distillation or debugging.
	type	boolean
	default	False
temperature	Temperature
	Temperature for sampling. The higher the temperature, the more random the completions.
	type	number
	default	1.0
epsilon_low	Epsilon Low
	Epsilon value for clipping.
	type	number
	default	0.2
epsilon_high	Epsilon High
	Upper-bound epsilon value for clipping. If not specified, it defaults to the same value as the lower-bound specified in argument epsilon. Paper DAPO recommends 0.28.
	type	number
	default	0.2
advantage_low	Advantage Low
	Lower-bound advantage value for clipping.
	type	number
	default	-5.0
advantage_high	Advantage High
	Upper-bound advantage value for clipping.
	type	number
	default	5.0
positive_nll_coef	Positive Nll Coef
	[Optional] Coefficient for Positive Example LM Loss. Set a positive value to enable; None disables.nRef: VAPO Sec. 4.3 (Positive Example LM Loss): https://arxiv.org/pdf/2504.05118
	default	None
	anyOf	type	number
		type	null
lower_bound_ratio	Lower Bound Ratio
	Lower-bound ratio for dual-clip.
	type	number
	default	3.0
loss_type	Loss Type
	The type of loss to use for GRPO training.
	type	string
	default	token-mean
unbiased_loss_max_tokens	Unbiased Loss Max Tokens
	Maximum number of tokens to use for unbiased loss introduced in Dr.GRPO. If set to None, will not use unbiased loss.Only available when loss_type is seq-mean-token-mean
	default	None
	anyOf	type	integer
		type	null
unbiased_advantage	Unbiased Advantage
	Whether to divide the advantage by the standard deviation of rewards.
	type	boolean
	default	False
overlong_reward	Configuration for overlong reward penalty. If enabled, the output will be penalized for responses that are too long.
	#/$defs/OverlongRewardConfig
kl_beta	Kl Beta
	KL coefficient. If 0.0, the reference model is not loaded, reducing memory usage and improving training speed, but may be numerically unstable for long training runs.
	type	number
	default	0.0
unbiased_kl_estimate	Unbiased Kl Estimate
	[Optional] Unbiased K3 with IS: D_KL ≈ E_{π_old}[ w · ( r − log r − 1 ) ], w=π_θ/π_old, r=π_ref/π_θ.nNote: This option is ignored when kl_beta is 0.0.nRef: DeepSeek-V3.2 Sec.3.1 (Unbiased KL Estimate): https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/main/assets/paper.pdf
	type	boolean
	default	False
off_policy_masking_delta	Off Policy Masking Delta
	Off-Policy Sequence Masking threshold δ (None disables). Per-sequence mask: M_i = 0 if Â_i < 0 and (1/|o_i|)∑_t log[π_old(o_{i,t}|·)/π_θ(o_{i,t}|·)] > δ; else M_i = 1.Ref: DeepSeek-V3.2 Sec.3.1 Off-Policy Sequence Masking : https://huggingface.co/deepseek-ai/DeepSeek-V3.2/resolve/main/assets/paper.pdf.
	default	None
	anyOf	type	number
		type	null
aipo_rho	Aipo Rho
	Rho value for AIPO (Asynchronous Importance weighted Policy Optimization). The clipping constant of the importance sampling ratio, suggest [2,10]. reference: https://arxiv.org/pdf/2505.24034
	default	None
	anyOf	type	number
		type	null
mu_iterations	Mu Iterations
	Number of iterations per batch (denoted as μ in the algorithm).
	type	integer
	default	1
mini_batch	Mini Batch
	mini-batch size for GRPO training. Mini-batch is used to split the batch per optimization into smaller batches to fit into GPU memory.
	type	integer
	default	2
batch_size_per_optimize	Batch Size Per Optimize
	batch size for each optimization in GRPO training. The batch in each training step is split into smaller batches which each performs one step optimization. If not set, it will be the same as the whole batch size per GPU for each training step.
	default	None
	anyOf	type	integer
		type	null
max_token_len_per_mini_batch	Max Token Len Per Mini Batch
	Maximum token length per mini batch. If set, dynamic mini-batch sizing will be applied based on this limit.
	default	None
	anyOf	type	integer
		type	null
entropy_coeff	Entropy Coeff
	Coefficient for entropy regularization.
	type	number
	default	0.0
allowed_outdated_steps	Allowed Outdated Steps
	Allowed outdated-async steps for rollout engine. If the number of left uncompleted rollout samples is larger than the (allowed_outdated_steps + 1) * n_policy_replicas * train_batch_per_replica, then rollout engine traffic will be throttled.
	type	integer
	default	4
on_policy	On Policy
	Enable fully synchronized (on-policy) rollout. If set to True, the rollout engine will wait until the expected weight version is updated before next generation starts.
	type	boolean
	default	False
outdated_rollout_fetch_batch_size	Outdated Rollout Fetch Batch Size
	Number of outdated rollouts to fetch. If set to 0, the rollout engine will stop generating rollouts if the weight is outdated.
	type	integer
	default	0
min_filter_prefix_tokens	Min Filter Prefix Tokens
	Minimum number of tokens to filter the prefix tokens for the rollouts inside the same group. If the number of tokens is larger than the min_filter_prefix_tokens, the rollouts with the same prefix but different rewards will be filtered out in loss calculation.
	default	None
	anyOf	type	integer
		type	null
max_retry_for_on_policy	Max Retry For On Policy
	Maximum number of retries for on-policy rollout to have enough samples. If non-positive, will retry with no upper limit until enough samples are generated.
	type	integer
	default	-1
reference_reset_interval	Reference Reset Interval
	Interval to reset the reference model to the current model. If set to None or 0, the reference model will not be reset during training.
	default	None
	anyOf	type	integer
		type	null
reset_optimizer_with_reference	Reset Optimizer With Reference
	Whether to reset the optimizer state when the reference model is reset.
	type	boolean
	default	True
balance_dp_token	Balance Dp Token
	Whether to balance the number of tokens in each data parallel replica when calculating the loss.
	type	boolean
	default	False
use_decoupled_loss	Use Decoupled Loss
	Whether to use decoupled loss. A decoupled loss separates the optimization of the behavior policy and the target policy, which can help to reduce the variance of the gradient estimate.
	type	boolean
	default	False
behav_imp_weight_cap	Behav Imp Weight Cap
	Clipping cap for behavior importance weights. Useful when decoupled loss is used to avoid large variance.
	default	None
	anyOf	type	number
		type	null
rollout_as_token_ids	Rollout As Token Ids
	Whether to use token ids for rollouts instead of text. This can save tokenization time during rollout generation.
	type	boolean
	default	False
collect_rollout_logprobs	Collect Rollout Logprobs
	Whether to collect logprobs for rollouts instead of text. This can save logprob calculation time during rollout generation.
	type	boolean
	default	False
use_rollout_logprobs_for_loss	Use Rollout Logprobs For Loss
	Whether to use collected logprobs from rollouts for loss calculation. This is an alternative to calculating logprobs during training as old logprobs for importance sampling.
	type	boolean
	default	False
LoggingConfig	LoggingConfig
	type	object
	properties
	logger	Logger
		List of loggers to use, e.g., [‘console’, ‘wandb’]
		type	array
		items	type	string
	project_name	Project Name
		Wandb project name for logging. If set, the training will be logged to this project.
		type	string
		default	cosmos_rl
	group_name	Group Name
		Wandb group name for logging. If set, the training will be logged to this group.
		default	None
		anyOf	type	string
			type	null
	experiment_name	Experiment Name
		A short display name for this run. If not set, will use the output_dir as the experiment name.
		default	None
		anyOf	type	string
			type	null
LoraConfig	LoraConfig
	type	object
	properties
	r	R
		LoRA rank
		type	integer
		default	8
	lora_names	Lora Names
		A List of name for the LoRA adapters. If multiple names are provided, then multiple LoRA adapters will be created and trained simultaneously.
		type	array
		default	default
		items	type	string
	lora_path	Lora Path
		Path to pre-trained LoRA weights
		default	None
		anyOf	type	string
			type	null
	lora_alpha	Lora Alpha
		LoRA alpha
		type	number
		default	8.0
	lora_dropout	Lora Dropout
		LoRA dropout
		type	number
		default	0.0
	target_modules	Target Modules
		LoRA target modules, can be a list of strings or ‘all-linear’
		default	None
		anyOf	type	array
			items	type	string
			type	string
	use_rslora	Use Rslora
		When set to True, uses [Rank-Stabilized LoRA](https://huggingface.co/papers/2312.03732) which sets the adapter scaling factor to lora_alpha/math.sqrt(r), since it was proven to work better. Otherwise, it will use the original default value of lora_alpha/r.
		type	boolean
		default	False
	modules_to_save	Modules To Save
		List of modules apart from LoRA layers to be set as trainable and saved in the final checkpoint.
		default	None
		anyOf	type	array
			items	type	string
			type	null
	alpha_pattern	Alpha Pattern
		Per-module overrides for lora_alpha. Keys are regex patterns; evaluated in insertion order, first match wins.
		default	None
		anyOf	type	object
			additionalProperties	type	number
			type	null
	r_pattern	R Pattern
		Per-module overrides for LoRA rank r. Keys are regex patterns; evaluated in insertion order, first match wins.
		default	None
		anyOf	type	object
			additionalProperties	type	integer
			type	null
	init_lora_weights	Init Lora Weights
		How to initialize the weights of the adapter layers.Passing True (default) results in the default initialization from the reference implementation from Microsoft, with the LoRA B weight being set to 0. This means that without further training, the LoRA adapter will be a no-op.Setting the initialization to False leads to random initialization of LoRA A and B, meaning that LoRA is not a no-op before training; this setting is intended for debugging purposes.Passing ‘gaussian’ results in Gaussian initialization scaled by the LoRA rank for linear and layers. Pass ‘loftq’ to use LoftQ initialization. Passing ‘eva’ results in a data-driven initialization of Explained Variance Adaptation.EVA initializes LoRA based on the SVD of layer input activations and achieves SOTA performance due to its ability to adapt to the finetuning data. Pass ‘olora’ to use OLoRA initialization. Passing ‘pissa’ results in the initialization of https://huggingface.co/papers/2404.02948
		default	True
		anyOf	type	boolean
			type	string
			enum	gaussian, eva, olora, pissa, pissa_niter_[number of iters]
MultiTurnRolloutConfig	MultiTurnRolloutConfig
	type	object
	properties
	enable	Enable
		Whether to enable multi-turn rollout.
		type	boolean
		default	False
	enable_tools	Enable Tools
		Whether to enable tools in multi-turn rollout.
		type	boolean
		default	False
	enable_thinking	Enable Thinking
		Whether to enable thinking in multi-turn rollout.
		type	boolean
		default	False
	custom_chat_template_path	Custom Chat Template Path
		The path to the custom chat template in chat.
		default	None
		anyOf	type	string
			type	null
	max_assistant_turns	Max Assistant Turns
		Max assistant turn count for multi-turn rollout.
		type	integer
		default	5
	add_generation_prompt	Add Generation Prompt
		Whether to add generation prompt in multi-turn rollout.
		type	boolean
		default	True
	continue_final_message	Continue Final Message
		Whether to continue the final message in multi-turn rollout.
		type	boolean
		default	False
OverlongRewardConfig	OverlongRewardConfig
	type	object
	properties
	enable_overlong_penalty	Enable Overlong Penalty
		Enable overlong penalty for the model. If set to True, the output will be penalized for responses that are too long.
		type	boolean
		default	False
	buffer_length	Buffer Length
		Length of the buffer for overlong penalty. If the response length exceeds this value, the output will be penalized.
		type	integer
		default	4096
	penalty_factor	Penalty Factor
		Penalty factor for overlong penalty. The penalty increases linearly with the length of the response exceeding the buffer length from 0 to the penalty_factor.
		type	number
		default	1.0
ParallelismConfig	ParallelismConfig
	type	object
	properties
	n_init_replicas	N Init Replicas
		Number of initial replicas to be created
		type	integer
		default	1
	tp_size	Tp Size
		Tensor parallelism size
		type	integer
		default	2
	cp_size	Cp Size
		Context parallelism size
		type	integer
		default	1
	ep_size	Ep Size
		Expert parallelism size
		type	integer
		default	1
	dp_shard_size	Dp Shard Size
		Data Parallelism size in sharded mode
		type	integer
		default	1
	pp_size	Pp Size
		Pipeline parallelism size
		type	integer
		default	1
	pp_dynamic_shape	Pp Dynamic Shape
		Pipeline parallelism dynamic shape
		type	boolean
		default	False
	pp_micro_batch_size	Pp Micro Batch Size
		Pipeline parallelism micro batch size, n_micro_batch = batch_size / pp_micro_batch_size, which must be divisible by pp stages
		type	integer
		default	1
	dp_replicate_size	Dp Replicate Size
		Data Parallelism size in replica mode. Only configurable in SFT type job, must be 1 in GRPO type job for dynamic scaling support purpose.
		type	integer
		default	1
PolicyConfig	PolicyConfig
	type	object
	properties
	parallelism	#/$defs/ParallelismConfig
	diffusers	anyOf	#/$defs/DiffusersConfig
			type	null
	is_diffusers	Is Diffusers
		Whether this model is diffusers or not
		type	boolean
		default	False
	model_name_or_path	Model Name Or Path
		The model name or path, compatible with huggingface model name or local path
		type	string
		default	Qwen/Qwen2.5-VL-7B-Instruct
	model_revision	Model Revision
		The revision of the model to use
		default	None
		anyOf	type	string
			type	null
	model_max_length	Model Max Length
		The maximum length for training, longer than this will be ignored for training stability
		type	integer
		default	4096
	model_gradient_checkpointing	Model Gradient Checkpointing
		Whether to use gradient checkpointing
		type	boolean
		default	True
	lora	LoRA configuration
		default	None
		anyOf	#/$defs/LoraConfig
			type	null
	trainable_map	Trainable Map
		Mapping of name -> bool. Keys can either be: - exact parameter names (from model.named_parameters()) - exact module paths (from model.named_modules())
		default	None
		anyOf	type	object
			additionalProperties	type	boolean
			type	null
	freeze_pattern	Freeze Pattern
		Pattern-based configuration to freeze parts of the model. A list of regex patterns that match against parameter names; matched parameters will be frozen (requires_grad=False). Example: freeze_pattern = [‘^visual..*’] freezes all visual components; freeze_pattern = [‘^model.layers.[0-9]+.’] freezes layers 0-9.
		default	None
		anyOf	type	array
			items	type	string
			type	null
	enable_liger_kernel	Enable Liger Kernel
		Whether to use liger kernel.
		type	boolean
		default	False
ProfilerConfig	ProfilerConfig
	type	object
	properties
	enable_profiler	Enable Profiler
		Enable profiler for training
		type	boolean
		default	False
	enable_nsys	Enable Nsys
		Enable nsys for training
		type	boolean
		default	False
	sub_profiler_config	Sub profiler config
		#/$defs/SubProfilerConfig
RemoteRewardConfig	RemoteRewardConfig
	type	object
	properties
	enabled	Enabled
		type	boolean
		default	True
	score_key	Score Key
		type	string
		default	overall_reward
	scale	Scale
		type	number
		default	1.0
	reward_fn	Reward Fn
		Dictionary of reward functions and their weights for remote reward calculation.
		type	object
		additionalProperties	type	number
	reward_clip_min	Reward Clip Min
		type	number
		default	-5.0
	reward_clip_max	Reward Clip Max
		type	number
		default	5.0
RolloutAsyncConfig	RolloutAsyncConfig
	type	object
	properties
	max_concurrent_requests	Max Concurrent Requests
		Maximum number of concurrent requests for rollout engine.
		type	integer
		default	10
RolloutConfig	RolloutConfig
	type	object
	properties
	parallelism	#/$defs/RolloutParallelismConfig
	enforce_eager	Enforce Eager
		Whether to enable eager execution for vLLM.
		type	boolean
		default	True
	include_stop_str_in_output	Include Stop Str In Output
		Whether to include stop string in output.
		type	boolean
		default	False
	gpu_memory_utilization	Gpu Memory Utilization
		GPU memory utilization factor for rollout backend.
		type	number
		default	0.8
	enable_chunked_prefill	Enable Chunked Prefill
		Whether to enable chunked prefill for vLLM.
		type	boolean
		default	False
	max_response_length	Max Response Length
		Max output length of rollout generation.
		type	integer
		default	2048
	n_generation	N Generation
		n parameter same like what in OpenAI chat API.
		type	integer
		default	16
	n_generation_to_batch	N Generation To Batch
		Whether to treat n_generation as batch dimension in rollout generation.
		type	boolean
		default	False
	batch_size	Batch Size
		Batch size for rollout.
		type	integer
		default	1
	quantization	Quantization
		Quantization in vllm rollout generation.
		type	string
		default	none
	seed	Seed
		random seed for rollout.
		default	None
		anyOf	type	integer
			type	null
	sampling_config	#/$defs/SamplingConfig
	vllm_use_flashinfer	Vllm Use Flashinfer
		Use flashinfer for vllm rollout.
		type	boolean
		default	False
	backend	Backend
		Backend for rollout. Currently support vllm, vllm_async and trtllm, and other custom backends.
		type	string
		default	vllm
	multi_turn_config	Configuration for multi-turn rollout.
		#/$defs/MultiTurnRolloutConfig
	mode	Mode
		Rollout mode, could be ‘sync’ or ‘async’.
		type	string
		default	sync
	async_config	Configuration for async rollout.
		#/$defs/RolloutAsyncConfig
RolloutParallelismConfig	RolloutParallelismConfig
	type	object
	properties
	n_init_replicas	N Init Replicas
		Number of initial replicas to be created
		type	integer
		default	1
	tp_size	Tp Size
		Tensor parallelism size
		type	integer
		default	2
	cp_size	Cp Size
		Context parallelism size
		type	integer
		default	1
	ep_size	Ep Size
		Expert parallelism size
		type	integer
		default	1
	dp_shard_size	Dp Shard Size
		Data Parallelism size in sharded mode
		type	integer
		default	1
	pp_size	Pp Size
		Pipeline parallelism size
		type	integer
		default	1
	pp_dynamic_shape	Pp Dynamic Shape
		Pipeline parallelism dynamic shape
		type	boolean
		default	False
	pp_micro_batch_size	Pp Micro Batch Size
		Pipeline parallelism micro batch size, n_micro_batch = batch_size / pp_micro_batch_size, which must be divisible by pp stages
		type	integer
		default	1
	dp_replicate_size	Dp Replicate Size
		Data Parallelism size in replica mode. Only configurable in SFT type job, must be 1 in GRPO type job for dynamic scaling support purpose.
		type	integer
		default	1
SFTDataConfig	SFTDataConfig
	type	object
	properties
	type	Type
		type	string
		const	sft
	trainer_type	Trainer Type
		Type of the trainer for SFT.
		type	string
		default	sft
	dataset	Dataset configuration for SFT training. It includes dataset name, subset, revision, train split, and test split.
		#/$defs/DatasetConfig
	mini_batch	Mini Batch
		mini-batch size for training.
		type	integer
		default	2
	dataloader_shuffle	Dataloader Shuffle
		Shuffle the dataloader. If False, the dataloader will be used in the order it is loaded.
		type	boolean
		default	True
	dataloader_seed	Dataloader Seed
		random seed for dataloader shuffling
		type	integer
		default	0
	enable_dataset_cache	Enable Dataset Cache
		Enable dataset cache process results, maybe accelerate the dataset loading
		type	boolean
		default	False
	dataloader_num_workers	Dataloader Num Workers
		Number of subprocess to use for data loading
		type	integer
		default	0
	dataloader_prefetch_factor	Dataloader Prefetch Factor
		Number of batches loaded in advance by each worker.
		default	None
		anyOf	type	integer
			type	null
	dataloader_drop_last	Dataloader Drop Last
		Whether to drop the last batch of the dataloader if it is not complete.
		type	boolean
		default	True
	data_dispatch_as_rank_in_mesh	Data Dispatch As Rank In Mesh
		Whether to dispatch data according to rank in global mesh. If True, each rank will get its specific data shard based on its rank in the global mesh.
		type	boolean
		default	False
	conversation_column_name	Conversation Column Name
		Column name for formated conversation json
		type	string
		default	conversations
	system_prompt	System Prompt
		System prompt for the model, which will be prepended to the prompt
		type	string
		default
	balance_dp_token	Balance Dp Token
		Whether to balance the number of tokens in each data parallel replica when calculating the loss.
		type	boolean
		default	True
SampleConfig	SampleConfig
	type	object
	properties
	num_steps	Num Steps
		Number of sampler inference steps for training
		type	integer
		default	40
	eval_num_steps	Eval Num Steps
		Number of sampler inference steps for evaluation
		type	integer
		default	40
	guidance_scale	Guidance Scale
		Classifier-free guidance weight
		type	number
		default	4.5
	global_std	Global Std
		Whether to use all samples in a batch to compute std
		type	boolean
		default	True
	noise_level	Noise Level
		Noise level for sampling
		type	number
		default	1.0
	deterministic_sampling	Deterministic Sampling
		Whether to use deterministic sampling
		type	boolean
		default	False
	solver	Solver
		Sampler solver to be used
		type	string
		default	dpm2
SamplingConfig	SamplingConfig
	type	object
	properties
	temperature	Temperature
		Temperature for sampling.
		type	number
		default	1.0
	top_p	Top P
		Top-p for sampling.
		type	number
		default	1.0
	top_k	Top K
		Top-k for sampling.
		type	integer
		default	-1
	repetition_penalty	Repetition Penalty
		Repetition penalty for sampling.
		type	number
		default	1.0
	use_flashinfer	Use Flashinfer
		Use flashinfer for sampling.
		type	boolean
		default	False
SubProfilerConfig	SubProfilerConfig
	type	object
	properties
	do_profile	Do Profile
		Whether to profile, only used in runtime.
		type	boolean
		default	False
	active_steps	Active Steps
		Number of active steps
		type	integer
		default	1
	warmup_steps	Warmup Steps
		Number of warmup steps
		type	integer
		default	1
	wait_steps	Wait Steps
		Number of wait steps
		type	integer
		default	1
	rank_filter	Rank Filter
		Rank filter
		type	array
		items	type	integer
	record_shape	Record Shape
		Whether to record shape
		type	boolean
		default	False
	profile_memory	Profile Memory
		Whether to profile memory
		type	boolean
		default	False
	with_stack	With Stack
		Whether to profile stack
		type	boolean
		default	False
	with_modules	With Modules
		Whether to profile modules
		type	boolean
		default	False
TokenizerConfig	TokenizerConfig
	type	object
	properties
	chunk_duration	Chunk Duration
		type	integer
		default	81
	load_mean_std	Load Mean Std
		type	boolean
		default	False
	compile_encode	Compile Encode
		type	boolean
		default	False
	temporal_window	Temporal Window
		type	integer
		default	16
TrainingConfig	TrainingConfig
	type	object
	properties
	train_policy	Train Policy
		default	type	grpo
			trainer_type	grpo
			variant	grpo
			dataset	local_dir
				name
				revision
				split
				subset
				test_size	None
			dataloader_shuffle	True
			dataloader_seed	0
			data_dispatch_as_rank_in_mesh	False
			enable_dataset_cache	False
			dataloader_num_workers	0
			dataloader_prefetch_factor	None
			dataloader_batch_size	1
			prompt_column_name
			response_column_name
			reward_function	single_choice	1.0
			use_remote_reward	False
			remote_reward	enabled	True
				reward_clip_max	5.0
				reward_clip_min	-5.0
				reward_fn	dance_grpo	1.0
				scale	1.0
				score_key	overall_reward
			filter_reward_metric
			bypass_reward	False
			temperature	1.0
			epsilon_low	0.2
			epsilon_high	0.2
			advantage_low	-5.0
			advantage_high	5.0
			positive_nll_coef	None
			lower_bound_ratio	3.0
			loss_type	token-mean
			unbiased_loss_max_tokens	None
			unbiased_advantage	False
			overlong_reward	buffer_length	4096
				enable_overlong_penalty	False
				penalty_factor	1.0
			kl_beta	0.0
			unbiased_kl_estimate	False
			off_policy_masking_delta	None
			aipo_rho	None
			mu_iterations	1
			mini_batch	2
			batch_size_per_optimize	None
			max_token_len_per_mini_batch	None
			entropy_coeff	0.0
			allowed_outdated_steps	4
			on_policy	False
			outdated_rollout_fetch_batch_size	0
			min_filter_prefix_tokens	None
			max_retry_for_on_policy	-1
			reference_reset_interval	None
			reset_optimizer_with_reference	True
			balance_dp_token	False
			use_decoupled_loss	False
			behav_imp_weight_cap	None
rollout_as_token_ids			False
collect_rollout_logprobs			False
use_rollout_logprobs_for_loss			False
oneOf	#/$defs/SFTDataConfig
	#/$defs/GrpoConfig
optm_name	Optm Name
	Optimizer name
	type	string
	default	AdamW
optm_lr	Optm Lr
	Learning rate for optimizer, can be a float or a list of floats for multiple optimizers
	default	1e-06
	anyOf	type	number
		type	array
		items	type	number
optm_impl	Optm Impl
	Implementation type for optimizer. More info: https://pytorch.org/docs/stable/optim.html, can be a list of strings for multiple optimizers
	default	fused
	anyOf	type	string
		type	array
		items	type	string
optm_weight_decay	Optm Weight Decay
	Weight decay for optimizer
	type	number
	default	0.01
optm_betas	Optm Betas
	Betas for optimizer
	type	array
	default	0.9
		0.999
	maxItems	2
	minItems	2
optm_warmup_steps	Optm Warmup Steps
	Warmup steps for optimizer, can be an integer or a float, if it is a float and range in [0.0, 1.0], it will be multiplied by the total steps
	default	20
	anyOf	type	integer
		type	number
optm_decay_ratio	Optm Decay Ratio
	Ratio of total steps for decay, range in [0.0, 1.0], 0 means no decay.
	default	None
	anyOf	type	number
		type	null
optm_decay_type	Optm Decay Type
	Type of decay for optimizer
	default	None
	anyOf	type	string
		type	null
optm_min_lr_factor	Optm Min Lr Factor
	Minimum lr factor for optimizer, range in [0.0, 1.0]
	type	number
	default	0.0
optm_grad_norm_clip	Optm Grad Norm Clip
	Gradient norm clip for optimizer
	type	number
	default	1.0
ema_enable	Ema Enable
	Whether to enable EMA for model parameters. Only support diffusers models for now.
	type	boolean
	default	False
ema_decay	Ema Decay
	Decay rate for EMA
	type	number
	default	0.9999
ema_update_step_interval	Ema Update Step Interval
	Interval steps to update EMA parameters, 0 means update every step
	type	integer
	default	0
master_dtype	Master Dtype
	The master weight data type for optimizers, is orthognal to param_dtype. Should be high precision for convergence consideration
	type	string
	default	float32
param_dtype	Param Dtype
	The data type for forward/backward. Outside forward/backward, params are in master_dtype
	type	string
	default	bfloat16
transfer_dtype	Transfer Dtype
	The data type for transfer parameters between Policy and Rollout.
	type	string
	default	None
logprob_dtype	Logprob Dtype
	The data type for logprobs calculation.
	type	string
	default	float32
fsdp_reduce_dtype	Fsdp Reduce Dtype
	The data type for reduction in FSDP
	type	string
	default	float32
fsdp_offload	Fsdp Offload
	Whether to offload the model to CPU if using FSDP
	type	boolean
	default	False
fsdp_reshard_after_forward	Fsdp Reshard After Forward
	Reshard the param after forward pass in FSDP
	type	string
	default	default
train_batch_per_replica	Train Batch Per Replica
	The batch size for training per iteration in one replica, this is the local batch size for each gradient accumulation step
	type	integer
	default	8
fp8	#/$defs/FP8Config
fp4	#/$defs/FP4Config
ckpt	#/$defs/CheckpointConfig
resume	Resume
	Resume training from a checkpoint. If True, will resume from the latest checkpoint of the output_dir. If a string, will resume from the specified checkpoint path.
	default	False
	anyOf	type	boolean
		type	string
epoch	Epoch
	Number of epochs for training
	type	integer
	default	1
output_dir	Output Dir
	Output directory
	type	string
	default	./outputs
timestamp	Timestamp
	Timestamp for the output directory and wandb ID, if not set, will be generated automatically
	type	string
	default
epsilon	Epsilon
	Epsilon for optimizer
	type	number
	default	1e-06
async_tp_enabled	Async Tp Enabled
	Whether to use async tensor parallelism
	type	boolean
	default	False
compile	Compile
	Whether to use torch.compile
	type	boolean
	default	True
sync_weight_interval	Sync Weight Interval
	The interval of train step for synchronizing weights between replicas.
	type	integer
	default	1
deterministic	Deterministic
	Whether to use deterministic training. If set to True, will use deterministic training, which is expected to be slower.
	type	boolean
	default	False
activation_offload	Activation Offload
	Whether to use activation offload
	type	boolean
	default	False
fa_version	Fa Version
	FlashAttention version to use. If None, will use the default version.
	default	None
	anyOf	type	integer
		type	null
seed	Seed
	Random seed for training. If deterministic is set to True, will by default be set to 42.
	default	None
	anyOf	type	integer
		type	null
local_dataset	Local Dataset
	Whether to use local dataset to query sample. If set to True, will use the local dataset.
	default	True
	anyOf	type	boolean
		type	null
force_use_hf	Force Use Hf
	Whether to force using Huggingface dataset even if local dataset is available.
	default	False
	anyOf	type	boolean
		type	null
non_text	Non Text
	Whether train in non-text mode. If set to True, the inputs and outputs are not pure text, but may contain other modalities like images, videos, tensors, etc.
	type	boolean
	default	False
max_num_steps	Max Num Steps
	Optional upper bound on total training steps. If set, training stops when either this step count or the epoch-based limit is reached (whichever comes first). Handy for quick smoke tests.
	default	None
	anyOf	type	integer
		type	null
sequence_packing	Sequence Packing
	Whether to enable sequence packing for training. If set to True, the input sequences will be packed into a single tensor for training.
	type	boolean
	default	False
VLAConfig	VLAConfig
	type	object
	properties
	vla_type	Vla Type
		VLA type, could be ‘openvla-oft’ or ‘openvla’
		type	string
		default	openvla-oft
	num_envs	Num Envs
		Number of environments to rollout.
		type	integer
		default	1
	use_proprio	Use Proprio
		Whether to use proprioceptive information.
		type	boolean
		default	False
	proprio_dim	Proprio Dim
		Dimension of proprioceptive information.
		type	integer
		default	7
	num_images_in_input	Num Images In Input
		Number of images in input.
		type	integer
		default	1
	training_chunk_size	Training Chunk Size
		Number of chunks to train in one iteration.
		type	integer
		default	16
	save_video	Save Video
		Whether to save video of validation rollout.
		type	boolean
		default	False
	continuous	Continuous
		Whether to enable continuous simulation + rollout.
		type	boolean
		default	False
	trace_verbosity	Trace Verbosity
		Verbosity level for tracing. 0=disabled, 1=validation only, 2=all.
		type	integer
		default	1
ValidationConfig	ValidationConfig
	type	object
	properties
	enable	Enable
		Enable validation during training.
		type	boolean
		default	False
	val_before_train	Val Before Train
		Enable validation before training starts (at step 0, after weight initialization).
		type	boolean
		default	False
	freq	Freq
		Validation frequency during training, in terms of training steps
		type	integer
		default	20
	batch_size	Batch Size
		Batch size for validation, will use the same batch size as training if not set.
		default	None
		anyOf	type	integer
			type	null
	dataset	Dataset configuration for validation. It includes dataset name, subset, revision and test split.
		#/$defs/DatasetConfig
	temperature	Temperature
		Temperature for sampling during validation.
		type	number
		default	0.0
	top_p	Top P
		Top-p for sampling during validation.
		default	None
		anyOf	type	number
			type	null
	top_k	Top K
		Top-k for sampling during validation.
		default	1
		anyOf	type	integer
			type	null
	repetition_penalty	Repetition Penalty
		Repetition penalty for sampling during validation.
		type	number
		default	1.0
	n_generation	N Generation
		n parameter same like what in OpenAI chat API for validation.
		type	integer
		default	1
	max_response_length	Max Response Length
		Max output length of rollout generation during validation.
		default	None
		anyOf	type	integer
			type	null
	reward_function	Reward Function
		Reward functions for the model. Currently support single_choice, boxed_math, and format. You can add weight to each reward function by passing a dict, e.g., {‘single_choice’: 0.9, ‘format’: 0.1}
		default
		anyOf	type	string
			type	array
			items	type	string
			type	object
			additionalProperties	type	number