From a fact in the data to a number on a leaderboard

From a fact in the data to a number on a leaderboard A workflow diagram generated by Archify. EX / Misconceptions about causes and results 02 / From a fact to a number (main) 03 / Where the measurement drifts EX / Misconceptions about confidence and measurement Bigger models fix it · long tail still fails · Misconceptions about causes and results Bigger models fix it long tail still fails Add correct data · new facts teach guessing · Misconceptions about causes and results Add correct data new facts teach guessing Says 90%, right 90% · verbalized is overconfident · Misconceptions about causes and results Says 90%, right 90% verbalized is overconfident Measures general skill · measures a narrow slice · Misconceptions about causes and results Measures general skill measures a narrow slice 1–2 points = better · often within noise · Misconceptions about causes and results 1–2 points = better often within noise Fact in the data · common or long-tail · From a fact to a number (main) Fact in the data common or long-tail Pretraining · learns the data distribution · From a fact to a number (main) Pretraining learns the data distribution Post-training · SFT, RLHF · From a fact to a number (main) Post-training SFT, RLHF Decode an answer · right, wrong or abstain · From a fact to a number (main) Decode an answer right, wrong or abstain Grading · answer key or LLM judge · From a fact to a number (main) Grading answer key or LLM judge Leaderboard number · one score, rarely a CI · From a fact to a number (main) Leaderboard number one score, rarely a CI Test set leaks in · contamination · Where the measurement drifts Test set leaks in contamination Confidence · logprob, verbal, consistency · Where the measurement drifts Confidence logprob, verbal, consistency Judge bias · position, length, self · Where the measurement drifts Judge bias position, length, self Not found means clean · tests are one-sided · Misconceptions about confidence and measurement Not found means clean tests are one-sided Confident means right · calibration is per group · Misconceptions about confidence and measurement Confident means right calibration is per group Judges are objective · they carry their own bias · Misconceptions about confidence and measurement Judges are objective they carry their own bias Elo is absolute · depends on prompts, process · Misconceptions about confidence and measurement Elo is absolute depends on prompts, process leaks in probabilities grades for humans Legend model step source of measurement bias measuring and grading data common misconception