Skip to main content

Crate eredu_evaluation

Crate eredu_evaluation 

Source
Expand description

Backend-neutral model evaluation drivers.

Structs§

CheckpointParityOptions
Configuration for text-checkpoint logit parity.
CheckpointParityReport
General parity report plus source artifact identities.
DistributionMetrics
Metrics comparing one candidate categorical distribution with a reference.
EvaluationEvidence
Portable evidence emitted by an evaluator, backend probe, or reference.
LatencySummary
Summary of repeated operation latencies.
LogitRowMetrics
Metrics for one vocabulary-logit row.
LogitTolerance
Thresholds specialized for rows of vocabulary logits.
NumericMetrics
Common metrics for floating-point vectors.
NumericTolerance
Absolute and relative thresholds for ordinary floating-point observations.
ObservationParity
Comparison result for one path.
ParityPolicy
Complete policy for comparing two observation sets.
ParityReport
Complete backend/reference parity report.
ParityRule
One selector-specific comparison rule.
PersonaPlexEvaluationOptions
Controls for one PersonaPlex comparison.
PersonaPlexEvaluationPaths
Paths consumed and produced by one PersonaPlex comparison.
RealtimeTrace
Completed portable outputs from one realtime request.

Enums§

CheckpointParityError
Invalid checkpoint evidence or parity policy.
DistributionError
Invalid categorical-distribution comparison.
EvidenceError
Invalid evaluation evidence.
ParityComparison
Comparison applied to selected observation paths.
ParityError
Invalid comparison policy or incompatible observations.
ParityMetrics
Metrics emitted for one compared observation.
RealtimeTraceError
Invalid portable realtime trace evidence.

Constants§

DEFAULT_SAMPLING_SEED
Default deterministic sampling seed.
DEFAULT_TEXT_PROMPT
Default PersonaPlex system instruction used by the evaluator.

Functions§

compare_checkpoint_artifacts
Compares two checkpoint-probe report/tensor pairs using general parity.
compare_distributions
Compares two complete categorical logit vectors.
compare_observations
Compares two portable observation sets under one policy.
encoded_audio_frames
Converts neutral [batch, codebooks, frames] integer tokens to frame inputs.
observe_f32_tensor
Materializes a neutral tensor as F32 evaluation evidence.
observe_i32_tensor
Materializes a neutral tensor as signed integer evaluation evidence.
observe_realtime_frame
Converts a completed portable realtime frame into general observations.
run_personaplex_quantization
Runs the complete PersonaPlex dense-versus-quantized evaluation.
run_realtime_trace
Executes portable encoded frames through any realtime backend.
summarize_latencies
Summarizes finite, nonnegative latency samples.