DalšíDalší průvodce
Statistical Significance in LLM Evals
Technický
Technický PRŮVODCE
A token logprob expresses the model’s conditional probability for a token in its generated sequence; it is a property of token choice, not automatically a calibrated probability that the full answer is correct.
Poslechněte si tento návodAI-vyprávěný · ~3 min
Logprobs can contribute to task-specific confidence estimates when validated against labeled outcomes.
A log probability is the logarithm of a probability. In generation, a token logprob describes how likely the model assigned a particular next token given the preceding context and model state. Some APIs can return logprobs for generated tokens and, in some endpoints, alternatives among the most likely tokens. Availability and limits depend on provider, endpoint, and model. These values can help inspect ambiguity or compare candidate labels in a constrained classification task. But natural language answers consist of multiple tokens, and tokenization, answer length, wording, and prompt design affect sequence scores. A high probability for each token does not mean the answer is factually correct; models may confidently generate common but false continuations. Summing or averaging token logprobs also needs careful normalization and validation. Confidence calibration asks whether examples assigned a confidence level are correct at a corresponding frequency. To use logprobs as a confidence feature, define the target outcome, gather representative labeled examples, derive a score, and evaluate calibration and error detection on held-out data. Possible measures include reliability diagrams, Brier score, expected calibration error, and selective risk at a chosen abstention threshold. Research on language-model calibration finds that token probabilities and verbalized confidence can behave differently across model families and tasks. Treat logprobs as one signal, not an answer-level truth meter. Use human review or a fallback for high-impact decisions, and check whether the required API actually exposes logprobs for the selected model.
A classifier compares logprobs for a fixed set of mutually exclusive labels, then calibrates scores on labeled data.
A developer inspects token alternatives to understand whether a constrained answer was ambiguous.
A team checks Brier score and expected calibration error on a held-out validation set.
A high-impact workflow routes low-confidence cases to human review rather than accepting token scores as truth.
Providers may expose richer token-level signals, while research continues to study semantic confidence and calibration across tasks. Better confidence systems will combine probabilities with task context, external evidence, and held-out calibration data. Token logprobs will remain most useful when the output space is constrained and outcomes can be labeled. Future applications should report error rates and calibration drift rather than showing an unvalidated confidence number to users. Calibration methods will need updates when the model, prompt, or input population changes.
Newsletter zdarma
Tři ověřené příběhy ze světa AI každý všední den ráno, psané srozumitelnou angličtinou. Navždy zdarma, bez reklam.
Jeden e-mail každý všední den. Odhlásit se můžete jedním kliknutím. Vaši adresu nikdy neprodáváme ani nesdílíme.
Otestujte se
Okamžitá zpětná vazba ke každé odpovědi a sdílitelný certifikát s ověřitelným ID po absolvování kurzu.
Podpořte bezplatnou AI vzdělávání. AI Understanding je nezisková organizace 501(c)(3) — žádné reklamy, žádná platba za přístup, nikdy. Darujte
A token logprob expresses the model’s conditional probability for a token in its generated sequence; it is a property of token choice, not automatically a calibrated probability that the full answer is correct. Logprobs can contribute to task-specific confidence estimates when validated against labeled outcomes.
Token likelihood and factual correctness are different quantities.
Sequence aggregation depends on how the text is tokenized and composed.
Calibration requires comparison between predicted confidence and empirical outcomes.
Calibration compares stated confidence levels with observed accuracy.
A bounded output space can make score construction more interpretable, though validation is still required.
Učte se dál
Pro toto téma bylo vybráno více průvodců
DalšíDalší průvodce
Statistical Significance in LLM Evals
Technický