Skip to main content

Metrics reference

Generated from every built-in metric op manifest, including supplemental manifests.

MetricDescriptionAvailability
agent.custom_judge_rubricJudges a SPAN or TRACE partition against a fully customer-authored Jinja2 prompt template and judge model, with no fixed rubric of its own.Available
agent.hallucinationMeasures how well agent text is grounded in real evidence, via LLM-judge scoring at either single-completion (span) or whole-trace granularity.Available
agent.step_accuracyMeasures trajectory adherence — whether the agent's steps matched the expected sequence, including any unordered concurrent-step groups — via edit-distance alignment by default or strict position via config. Supports trace-level and session-level scoring.Available
agent.subagent_validityRule-based (no LLM judge) validity classification for every sub-agent (AGENT-kind) span's response - valid, empty, short, or error - reduced per partition to whether every sub-agent behaved validly this turn (TRACE) or the ratio of valid sub-agents across the session (SESSION).Available
agent.tool_selection_accuracyMeasures tool-selection coverage — how many of the required tools the agent called, anywhere in the trace or session, regardless of order or repeat count. A trace with no TOOL spans at all falls back to reading tool calls embedded in its LLM spans. Scored per trace or pooled across a session; there is no span-level scoring.Available
image_classification.accuracyThe fraction of images whose top-scoring predicted class matches the ground-truth label.Available
image_classification.confusion_matrixThe dataset-level table of true-label versus predicted-label counts, the basis for the per-class rate metrics.Available
image_classification.fnrThe average one-vs-rest false-negative rate across classes - how often true members of a class are missed.Available
image_classification.per_class_fnrFalse-negative rate computed separately for each class label, emitted as a per-class breakdown alongside the macro-average score.Available
image_classification.per_class_recallRecall computed separately for each class label, emitted as a per-class breakdown alongside the macro-average score.Available
image_classification.per_class_specificitySpecificity computed separately for each class label, emitted as a per-class breakdown alongside the macro-average score.Available
image_classification.recallThe average one-vs-rest recall across classes - how well the model finds each class's true members.Available
image_classification.roc_aucThe area under the ROC curve, measuring how well the model's class scores separate the positive class from the rest across all thresholds.Available
image_classification.specificityThe average one-vs-rest specificity across classes - how well the model rejects images that do not belong to each class.Available
instance_segmentation.mean_diceMean Dice overlap between matched predicted and ground-truth object instances.Available
instance_segmentation.mean_iouMean intersection-over-union between matched predicted and ground-truth object instances.Available
llm.answer_relevanceJudges whether the response stays on-topic and actually addresses what the question asked.Available
llm.biasJudges how free the response is of stereotyping or discrimination toward protected groups.Available
llm.bias_jailbreakMeasures how well the system resists technique-driven adversarial prompts that try to push it into producing biased content.Available
llm.bleuMeasures n-gram overlap between the response and the reference answer - how closely their wording matches.Available
llm.concisenessJudges how tightly the response sticks to the essentials, without padding, hedging, or filler.Available
llm.correctnessMeasures how well the facts in the response agree with the reference answer, in both accuracy and coverage.Available
llm.data_leakageMeasures how well the system resists adversarial attempts to extract private or internal data.Available
llm.data_leakage_jailbreakMeasures how well the system resists technique-driven adversarial attempts to extract private or internal data.Available
llm.decision_flipMeasures whether the system reaches a different decision for personas that differ along a single role or protected-attribute axis, when asked the same decision-eliciting question.Available
llm.exact_matchChecks whether the response is word-for-word identical to the reference answer.Available
llm.factual_consistencyMeasures whether the system agrees on facts, numbers, dates, and named entities when its query is reformulated, compared against its original answer.Available
llm.false_refusalMeasures how often the system refuses a benign question it should have answered — the over-cautious failure mode opposite to toxicity.Available
llm.group_interaction_biasMeasures how evenly the system treats users across different personas or group identities.Available
llm.helpfulnessJudges whether the response is both on-point and complete for the user's need.Available
llm.instruction_followingMeasures how reliably the system obeys explicit format and length instructions.Available
llm.ood_detectionMeasures how well the system detects and refuses queries that fall outside its intended use case.Available
llm.perturbation_robustnessMeasures how consistently the system stays correct across meaning-preserving variants of a question.Available
llm.perturbation_robustness_externalMeasures how consistently the system stays correct across meaning-preserving variants of a question, from a pre-scored dataset with off-platform SUT responses.Available
llm.policy_checkJudges whether the response complies with the deployer's content policy.Available
llm.rougeMeasures how much of the reference answer's content the response covers, using longest-common-subsequence overlap.Available
llm.sacrebleuMeasures response-vs-reference wording match on the standardised sacreBLEU scale.Available
llm.safety_consistencyMeasures whether the system's safety/policy posture changes when its query is adversarially reframed, compared against its original answer.Available
llm.semantic_consistencyMeasures whether the system reaches the same conclusion or stance when its query is paraphrased, compared against its original answer.Available
llm.toxicityJudges how free the response is of insults, hostility, or hateful language.Available
llm.toxicity_jailbreakMeasures how well the system resists adversarial prompts that try to push it into producing toxic content.Available
llm.toxicity_jailbreak_externalMeasures how well the system resists adversarial prompts that try to push it into producing toxic content.Available
llm.toxicity_robustnessMeasures how consistently the system stays non-toxic and de-escalates across many toxic-request variants.Available
object_detection.average_precisionMean Average Precision (mAP) over the whole dataset, summarising detection quality across a sweep of overlap thresholds.Available
object_detection.average_recallMean Average Recall (mAR) over the whole dataset, summarising how many ground-truth objects the detector finds across a sweep of overlap thresholds.Available
object_detection.bbox_area_mismatchHow closely the model's predicted box sizes match the true object sizes, per class.Available
object_detection.confusion_matrixHow often real objects are found and given the correct class label, reported as a recall score plus a detection confusion-matrix artifact.Available
object_detection.f1The harmonic mean of precision and recall, balancing false positives against missed objects.Available
object_detection.false_positive_spatial_heatmapCounts spatial locations of false-positive detections across an object-detection run and returns a 32x32 density grid plus per-class false-positive counts; does not produce a scalar score.Available
object_detection.iou_distributionHow tightly the model's boxes overlap the real objects, measured by mean intersection-over-union.Available
object_detection.missed_detection_biasHow evenly the model finds objects across classes, exposing classes that are disproportionately missed.Available
object_detection.missed_detection_rateHow often real objects go undetected - the complement of recall.Available
object_detection.precisionThe fraction of the model's detections that are correct.Available
object_detection.recallThe fraction of real objects the model successfully detects.Available
object_detection.size_wise_performanceSmall-object F1 degradation: how far the detector's F1 on small objects falls short of cv-eval's all-bands reference F1, measured at object-box grain.Available
object_detection.specificityHow reliably the model avoids raising detections for object classes that are not present in the image.Available
rag.content_biasMeasures how neutral and factual the system stays when a question is rephrased in a biased, one-sided way.Available
rag.context_precisionMeasures how much of the retrieved material is actually relevant to the answer, rewarding relevant chunks that are ranked near the top.Available
rag.context_recallMeasures how fully the retrieved material covers the evidence needed to support the reference answer.Available
rag.faithfulnessMeasures how well the answer sticks to the retrieved sources, penalising claims the context does not support.Available
rag.groundednessMeasures how well the answer's claims are supported by the retrieved evidence rather than outside knowledge or speculation.Available
rag.hit_at_kChecks whether the correct supporting evidence appears anywhere in the top-k retrieved results.Available
rag.retrieval_correctnessMeasures how well the retrieved results match the expected evidence, combining ranking quality (precision) with coverage (recall).Available
rag.retrieval_relevanceMeasures how relevant the retrieved material is to the question, judged independently of the answer.Available
semantic_segmentation.label_coverage_checkFraction of semantic-segmentation images with labels.Available
semantic_segmentation.label_format_checkFraction of semantic-segmentation labels with valid mask format.Available
semantic_segmentation.mean_diceMean Dice overlap between predicted and ground-truth semantic masks.Available
semantic_segmentation.mean_iouMean intersection-over-union between predicted and ground-truth semantic masks.Available
semantic_segmentation.per_sample_diceDice overlap for each image after semantic masks are decoded to dense arrays.Available
semantic_segmentation.pixel_accuracyPer-class share of ground-truth pixels the predicted mask labels correctly.Available
semantic_segmentation.prediction_coverage_checkFraction of semantic-segmentation images with predictions.Available