Değerlendirme

Discipline

1. Offline EvaluationHeld-out discipline and proxies: what offline numbers claim, and what they cannot.1–2 sa
2. Statistical Significance and Confidence Intervals for EvaluationIs the improvement real? Bootstrap intervals, paired tests, and test-set size.1–2 sa
3. Online Experiments and A/B TestingMeasuring real impact in production: randomised experiments, guardrail metrics, and peeking.1–2 sa
4. Baselines and AblationsWhat a result is compared against: trivial baselines, strong baselines, and ablations that isolate each component's contribution.1–2 sa

Supervised Metrics

1. Regression MetricsMAE, MSE, R-squared: reading what each regression number rewards.1–2 sa
2. Classification MetricsAccuracy, precision, recall, F1: the trade behind the single number.1–2 sa
3. Confusion MatricesError structure, tabulated: reading where the model confuses what.1–2 sa
4. ROC and Precision-Recall CurvesSweeping the decision threshold: every threshold is a confusion matrix, every curve is all of them at once.1–2 sa
5. Ranking MetricsEvaluating ordered results: precision@k, recall@k, MRR, and NDCG.1–2 sa

Generation Eval

1. Generation MetricsBLEU, ROUGE, and their limits: overlap scores for generated text.1–2 sa
2. LLM EvaluationBenchmarks, contamination, and reading leaderboard numbers honestly.1–2 sa
3. Human EvaluationProtocols, agreement, and cost: when only people can judge.1–2 sa
4. Benchmark Contamination and ValidityWhen benchmarks stop measuring: test data in the training set, saturation, and construct validity.1–2 sa
5. LLM-as-a-JudgeUsing models to grade model outputs, and measuring the judge's agreement with humans and its biases.1–2 sa

Trust

1. Robustness EvaluationPerturbations, slices, and stress tests: eval that fights back.1–2 sa
2. Error Analysis and SlicingReliability plus autopsy: confidence honesty and systematic error reading.1–2 sa