Imp.Optimizer.GRPO.Callback (Imp v0.5.0)

Copy Markdown View Source

Stable, data-only callback identity for resumable GRPO jobs.

A callback binds a trusted, already-loaded module and function to a consumer-owned versioned id and JSON-safe configuration. Imp persists only that identity and the configuration digest; it never serializes executable code or reconstructs atoms from saved data.

Reward callbacks receive (example, prediction, config). Validation callbacks receive (program, dataset, context, config).

Summary

Types

kind()
t()

Functions

identity(callback)
invoke(callback, example, prediction)
invoke(callback, program, dataset, context)
new(kind, module, function, opts)
reward(module, function, opts \\ [])
validate(callback, expected_kind)
validation(module, function, opts \\ [])

Types

kind()

@type kind() :: :reward | :validation

t()

@type t() :: %Imp.Optimizer.GRPO.Callback{
  config: term(),
  config_sha256: String.t(),
  function: atom(),
  id: String.t(),
  kind: kind(),
  module: module()
}

Functions

identity(callback)

invoke(callback, example, prediction)

invoke(callback, program, dataset, context)

new(kind, module, function, opts)

reward(module, function, opts \\ [])

validate(callback, expected_kind)

validation(module, function, opts \\ [])