Imp.Training.FastSlow.AdvantageGroup (Imp v0.5.0)

Copy Markdown View Source

Question-level rollout group with normalized training advantages.

Summary

Functions

Returns a deterministic JSON-safe representation for provider training boundaries.

Builds one question-level advantage group across every prompt allocation.

Types

t()

@type t() :: %Imp.Training.FastSlow.AdvantageGroup{
  cycle: non_neg_integer(),
  epsilon: float(),
  id: String.t(),
  mean_reward: float(),
  members: [map()],
  problem_id: String.t(),
  prompt_count: pos_integer(),
  size: pos_integer(),
  std_reward: float()
}

Functions

dump(group)

@spec dump(t()) :: map()

Returns a deterministic JSON-safe representation for provider training boundaries.

new!(rollouts, group_id, cycle, group_size, prompt_count, epsilon \\ 1.0e-6)

Builds one question-level advantage group across every prompt allocation.