Skip to contents

Extract frozen writing features for warm-start prediction

Source: R/warm_start_features.R
extract_warm_start_features.Rd

Warm-start prediction uses text features to predict relative writing quality before collecting pairwise comparisons. This function extracts the frozen English writing features; it does not train a model or estimate BTL scores.

Usage

extract_warm_start_features(
  ids,
  texts,
  schema = "writing_features_v1",
  python = NULL
)

Arguments

ids

Unique, nonmissing character or finite numeric item IDs. IDs are returned as character strings; blank IDs are not allowed.

texts

A nonempty character vector of the same length as ids, without missing values. Empty strings are allowed. Text is never trimmed or normalized.

schema

The frozen schema identifier, currently "writing_features_v1".

python

Optional path to an existing Python interpreter. With NULL, use an existing environment selected through reticulate. Automatic environment creation is disabled. Conflicting interpreter selections require correction or a fresh R session; this function never switches an initialized interpreter.

Value

A tibble with character item_id and the 20 numeric features in schema order, in the requested ID order. The warm_start_schema attribute records the schema identifier. Document-level undefined values remain NA.

Details

Extraction requires optional reticulate and the audited Python 3.12.3 stack. See warm_start_python_status() and the installed setup instructions located by system.file("python", "README.md", package = "pairwiseLLM"). Software and resources must be installed explicitly before extraction. Package loading, schema inspection, and later prediction from precomputed features do not require Python. The tested environment is Linux x86_64; Windows and macOS have not been validated. Features are not evidence of predictive validity.

The default English spaCy model, resource contents and package versions are checked before extraction. Missing resources fail without downloads. Entropy is divided by all spaCy tokens, including punctuation and whitespace, rather than the filtered n_tokens feature. Undefined values and valid zeros follow warm_start_feature_schema(). Zero-vector coherence preserves upstream values and warnings. No feature is imputed or replaced with zero.

See also

warm_start_python_status(), warm_start_feature_schema(), fit_warm_start_model()

Other adaptive warm start: ensemble_warm_start_models(), fit_warm_start_model(), make_warm_start_cv_plan(), make_warm_start_prior(), pairwiseLLM_warm_model, predict.pairwiseLLM_warm_ensemble(), predict.pairwiseLLM_warm_model(), prepare_warm_start_model(), register_warm_start_model(), save_warm_start_model(), summary.pairwiseLLM_warm_ensemble(), summary.pairwiseLLM_warm_predictions(), warm_start_coefficients(), warm_start_feature_schema(), warm_start_python_status()

Examples

if (FALSE) { # \dontrun{
# Explicitly provision the documented environment first.
extract_warm_start_features(c("a", "b"), c("A short text.", "Another text."),
  python = "/path/to/venv/bin/python")
} # }