Predict relative quality from a portable warm-start model

Usage

# S3 method for class 'pairwiseLLM_warm_model'
predict(object, newdata, ...)

Arguments

object: A valid pairwiseLLM_warm_model object.
newdata: Precomputed feature rows with item_id, all original frozen schema columns, and matching warm_start_schema metadata. IDs must be unique, nonmissing and nonblank. Numeric IDs normalize to character. Extra columns are ignored; required columns cannot be omitted even if training removed them.
...: Reserved for future extensions; must be empty.

Value

A tibble in newdata row order with character item_id, numeric raw_prediction, and numeric calibrated_prediction. Raw predictions are on the within-task standardized outcome scale, not the original BT/BTL scale. Public fits apply the stored OOF calibration intercept and slope to raw values. For uncalibrated core fits, calibrated predictions are NA_real_, never identity-calibrated substitutes. Attributes warm_start_schema and warm_start_model record schema identity and model metadata (format version, task ID, outcome definition, and calibration status), respectively.

Details

Prediction applies training medians, centers, and sample SDs, followed by the stored intercept and coefficients. It does not recompute preprocessing, train a model, load glmnet, initialize Python, or check a Python environment. Predictions are not calibrated Bayesian prior means or prior standard deviations. See pairwiseLLM_warm_model for the portable model contract.

Examples

if (requireNamespace("glmnet", quietly = TRUE) &&
    requireNamespace("withr", quietly = TRUE)) {
  local({
    # Synthetic features illustrate the interface, not predictive validity.
    example_features <- function(seed) {
      withr::local_seed(seed)
      fields <- warm_start_feature_schema()$feature
      x <- as.data.frame(matrix(runif(15 * length(fields)), nrow = 15))
      names(x) <- fields
      x$n_tokens <- 11:25
      x$token_length_mean <- 2 + 10 * x$token_length_mean
      x$token_length_std <- 0.2 + x$token_length_std
      x$dale_chall_readability_score <- 5 + 20 * x$dale_chall_readability_score
      x <- data.frame(item_id = as.character(1:15), x)
      attr(x, "warm_start_schema") <- "writing_features_v1"
      x
    }
    features <- example_features(3103)
    theta <- 10 + 0.4 * features$n_tokens - 2 * features$token_length_mean
    # A small alpha grid keeps this example fast; the default has 41 values.
    model <- fit_warm_start_model(features$item_id, theta, "synthetic-a",
      features = features, alpha_grid = c(0, 1))
    predictions <- predict(model, features)
    head(predictions)
  })
}
#> # A tibble: 6 × 3
#>   item_id raw_prediction calibrated_prediction
#>   <chr>            <dbl>                 <dbl>
#> 1 1                0.529                 0.553
#> 2 2                0.188                 0.199
#> 3 3                0.125                 0.134
#> 4 4               -1.67                 -1.73 
#> 5 5                0.790                 0.825
#> 6 6               -1.76                 -1.82