Log probability
A log probability is the logarithm, usually the natural logarithm, of a probability assigned to an event, outcome, token, or sequence.
Build a Reasoning Model (From Scratch)
The logarithm of a model-assigned probability, used here to score how likely a generated token or whole sequence is under the current parameters.
- How it is used
- Scores and ranks candidate responses during self-refinement, self-consistency tie-breaking, rollout evaluation, and reinforcement-learning updates.
- How deep it goes
- Derives log-softmax indexing, sequence summation, answer-only averaging, and policy-loss integration, but does not develop broader probability theory.
- Examples
- Uses Berlin-versus-Bridge answers, seven-token answer averaging, rollout comparisons, and a seven-token synthetic vocabulary.
- What you can do after
- Apply it via PyTorch to compute token, average answer, and sequence log probabilities for scoring responses and building GRPO losses.
Read how Build a Reasoning Model (From Scratch) explains log probability →
Statistics Every Programmer Needs
Log-likelihood is the logarithm of the likelihood of observed data under a statistical model and its parameter values.
- How it is used
- Evaluates and compares how well logistic-regression or time-series parameters explain observed data.
- How deep it goes
- Defines the likelihood transformation and applies it to null-model comparisons, pseudo R², and information criteria without deriving estimation algorithms.
- Examples
- Uses binary-outcome logistic regression and fitted time-series models as concrete settings for interpreting model fit.
- What you can do after
- Evaluate model fit and compare fitted models using stored log-likelihoods, pseudo R², AIC, BIC, and HQIC.
Read how Statistics Every Programmer Needs explains log probability →
Machine Learning Algorithms in Depth
A log likelihood is the logarithm of a model likelihood for observed data, commonly optimized instead of the likelihood itself.
- How it is used
- Uses log likelihood as a classification training objective and as a score for detecting anomalous sensor inputs.
- How deep it goes
- Derives optimization objectives and gradients for classifiers, then applies Gaussian log likelihood in an implemented anomaly detector.
- Examples
- Works through logistic regression, Bernoulli naive Bayes document classification, and an LSTM-VAE detecting anomalies in sensor time series.
- What you can do after
- Build it from scratch as a logistic-regression loss, or apply it via Keras/TensorFlow to threshold anomalous sensor readings.
Read how Machine Learning Algorithms in Depth explains log probability →
Build AI-Enhanced Web Apps
A log probability is the logarithm of the probability assigned to an event, such as a language-model token.
- How it is used
- Uses token likelihoods to estimate response confidence and support grounding decisions.
- How deep it goes
- Explains per-token scores and integration patterns, but stops short of formal probability mathematics and standardized framework support.
- Examples
- Uses OpenAI configuration, token-score response data, and a confidence-threshold RAG conversation handler.
- What you can do after
- Apply it via an API configuration to inspect token confidence and trigger grounded-response fallbacks.
Read how Build AI-Enhanced Web Apps explains log probability →
Generative AI in Action
Logprobs are completion metadata containing the natural logarithms of the model probabilities assigned to candidate tokens.
- How it is used
- Exposes token-level alternatives and confidence signals for debugging prompts, tuning generation, and selecting among completions.
- How deep it goes
- Defines natural-log token scores and demonstrates response arrays, but does not derive the underlying mathematics.
- Examples
- Uses a white miniature poodle naming request, Cotton’s token scores, alternatives, and best_of-generated pet-service names.
- What you can do after
- Diagnose/debug with it by inspecting alternatives, tune randomness, assess uncertainty, and select higher-scoring completions.
Causal AI
A log probability is the natural logarithm of a probability value.
- How it is used
- Uses logarithmic probabilities when working computationally with canonical probability distributions.
- How deep it goes
- Defines the transformation, explains its computational motivation, and demonstrates conversion back, but does not derive numerical-stability details.
- Examples
- Demonstrates Pyro's Bernoulli(0.4), log_prob for outcome 1.0, and math.exp reconstruction.
- What you can do after
- Apply it via Pyro and Python by evaluating log_prob and exponentiating the result.
Math and Architectures of Deep Learning
The natural logarithm of a probability density or mass value, used to make products and tiny probabilities easier to compute with.
- How it is used
- Supports probability computation and parameter estimation by replacing difficult likelihood products with simpler additive log-likelihoods.
- How deep it goes
- Explains the product-to-sum identity and optimization rationale, then applies them to Gaussian estimation without deriving every distribution internally.
- Examples
- Uses exponential identities, Gaussian likelihood optimization, adult height samples, and PyTorch distribution calculations.
- What you can do after
- Apply it via PyTorch distributions and optimize Gaussian or other exponential-family likelihoods in logarithmic form.
Read how Math and Architectures of Deep Learning explains log probability →
Reinforcement Learning from Human Feedback
A log probability is the logarithm of the probability assigned to an event or token by a probabilistic model.
- How it is used
- Computes policy-update signals from token probabilities in reinforcement-learning language-model training.
- How deep it goes
- Explains token-level log-probability differences, summed completion likelihoods, and their effects under three advantage cases.
- Examples
- Uses positive, negative, and zero advantages alongside the pg_loss = -advantages * ratio implementation.
- What you can do after
- Apply it via a policy-gradient loss to reinforce, suppress, or preserve sampled completions.
Read how Reinforcement Learning from Human Feedback explains log probability →
Related concepts
- policy gradient methodused by · A policy gradient method is a reinforcement-learning technique that uses gradients of expected return to optimize the parameters of a policy.
- generationpart of · Generation is the computational production of outputs from supplied inputs or context, often by applying a model to produce predictions or new content.
- completions apipart of · A completions API is an interface for requesting language-model-generated text from supplied prompts or inputs and configuring generation behavior.
- McFadden's pseudo R-squaredused by · McFadden's pseudo R-squared is a likelihood-based goodness-of-fit measure comparing a fitted discrete-choice model with an intercept-only null model.
- policyuses · A policy is a rule, strategy, or mapping that governs or determines decisions, actions, permissions, or behavior under specified conditions.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.