lloyal::metrics Namespace Reference

Namespaces
namespace	detail

Classes
struct	BranchMetricsState
	Unified model + sampling perplexity tracker. More...
struct	PerplexityState
	Rolling NLL accumulator for perplexity computation. More...

Enumerations
enum class	Base { Nats , Bits }

Functions
float	model_surprisal (const float *logits, int n_vocab, int picked_id, Base base=Base::Nats)
float	model_entropy (const float *logits, int n_vocab, Base base=Base::Nats)
float	sampling_surprisal (const float *candidate_logits, const int32_t *candidate_ids, int n_candidates, int picked_id, Base base=Base::Nats)
	Compute sampling-level surprisal for picked token.
float	sampling_entropy (const float *candidate_logits, int n_candidates, Base base=Base::Nats)
	Compute sampling-level entropy of candidate distribution.

Enumeration Type Documentation

Base

enum class lloyal::metrics::Base

strong

Enumerator
Nats
Bits

Examples

/home/runner/work/liblloyal/liblloyal/include/lloyal/metrics.hpp.

Definition at line 44 of file metrics.hpp.

Function Documentation

model_entropy()

float lloyal::metrics::model_entropy ( const float *  logits, int  n_vocab, Base  base = Base::Nats  )

inline

Examples

/home/runner/work/liblloyal/liblloyal/include/lloyal/metrics.hpp.

Definition at line 180 of file metrics.hpp.

model_surprisal()

float lloyal::metrics::model_surprisal ( const float *  logits, int  n_vocab, int  picked_id, Base  base = Base::Nats  )

inline

Examples

/home/runner/work/liblloyal/liblloyal/include/lloyal/branch.hpp, and /home/runner/work/liblloyal/liblloyal/include/lloyal/metrics.hpp.

Definition at line 131 of file metrics.hpp.

sampling_entropy()

float lloyal::metrics::sampling_entropy ( const float *  candidate_logits, int  n_candidates, Base  base = Base::Nats  )

inline

Compute sampling-level entropy of candidate distribution.

Measures uncertainty within the filtered candidate set (after top-k/p/temperature). Use to monitor distribution health after grammar masks or constraints.

Parameters

candidate_logits	Logits of candidate tokens (post-filter)
n_candidates	Number of candidates
base	Nats (natural log) or Bits (log₂)

Returns

Entropy in nats or bits (≥0, Infinity if invalid)

Examples

/home/runner/work/liblloyal/liblloyal/include/lloyal/metrics.hpp.

Definition at line 272 of file metrics.hpp.

sampling_surprisal()

float lloyal::metrics::sampling_surprisal ( const float *  candidate_logits, const int32_t *  candidate_ids, int  n_candidates, int  picked_id, Base  base = Base::Nats  )

inline

Compute sampling-level surprisal for picked token.

Measures uncertainty within the filtered candidate set (after top-k/p/temperature). Lower than model surprisal if filters removed low-probability tokens.

Use to monitor runtime hazard when grammar/constraints narrow the distribution.

Parameters

candidate_logits	Logits of candidate tokens (post-filter)
candidate_ids	Token IDs of candidates
n_candidates	Number of candidates
picked_id	Token ID that was sampled
base	Nats (natural log) or Bits (log₂)

Returns

Surprisal in nats or bits (≥0, Infinity if invalid)

Examples

/home/runner/work/liblloyal/liblloyal/include/lloyal/branch.hpp, and /home/runner/work/liblloyal/liblloyal/include/lloyal/metrics.hpp.

Definition at line 226 of file metrics.hpp.