protomotions.agents.peft.sft_agent module

SFT agent for PEFT adapters on a frozen discrete-token GPC prior.

class protomotions.agents.peft.sft_agent.DiscretePriorPEFTSFTAgent(fabric, env, config, root_dir=None)

Bases: DiscretePriorPEFTSetupMixin, PretrainedModulesMixin, SupervisedAgent

Train a discrete-prior PEFT adapter with target-token supervision.

The model owns the expert labeling path through the frozen target encoder; the generic supervised loop stores those labels and applies the configured supervised loss during optimization.

__init__(fabric, env, config, root_dir=None)

property has_critic

Whether this agent has a value critic.

Base agents do not assume critic ownership; PPO-style subclasses define that contract explicitly from their model config. Algorithm code should branch on this property instead of probing for critic attributes.

create_model()

create_optimizers(model)

get_state_dict(state_dict)

Get complete state dictionary for checkpointing.

Collects all agent state including model weights, training progress, and normalization statistics into a single dictionary for saving.

Parameters:

state_dict – Existing state dict to update (typically empty dict).

Returns:

Updated state dictionary containing all agent state.