glin

Permanently delete a trained model and its saved files. This cannot be undone; confirm with the user before you call it. To replace a model with a new one, call `train_model` with overwrite=True instead.

Global view of what a trained model learned, exact rather than sampled from example predictions: every feature ranked by its mean absolute contribution to the log-odds (`features`, with main-effect vs interaction split), the top terms including pairwise interactions (`terms`, each with its share of the total), and, for binary targets, a `shape` per main effect showing which categories or value ranges push the prediction up or down. Use this to answer "what drives this model" instead of running `predict` on a handful of rows.

Return feature names, types, and target classes for a trained model.

Every EBM training hyperparameter `train_model` accepts in its `hyperparameters` argument: name, type, default, allowed range, and what raising/lowering it does. Fetch the `tune_model_hyperparameters` prompt for a search strategy built on these.

List all trained glin models available on this machine.

Predicted class and probabilities, plus the full glassbox audit: `base_logit` (the intercept), every contributing feature's exact score (sorted by magnitude), and an additivity check verifying the scores sum to the model's own predicted probabilities (`reconstructed_probability` vs `model_probability`, both keyed by class name). For binary targets scores and `base_logit` are log-odds of `score_class` (`target_classes[1]`; positive pushes toward it, negative toward the other class); multiclass reports per-class scores and a per-class `base_logit`. `warnings` names inputs the model never saw (unseen categories, numerics or dates outside the training range, unparseable numbers or dates); treat the prediction as less reliable when present. `missing_features` lists absent inputs, which are valid but worth mentioning. Pass a date column as its raw value (e.g. "2024-03-06"); glin derives the `<col>__<part>` terms shown in `contributions` itself.

Train an EBM classifier from a CSV or Excel file and save it as model_name. This is also how to retrain an existing model: pass overwrite=True to replace it with a freshly trained model (e.g. on updated data) -- otherwise a pre-existing model_name is rejected so it isn't silently clobbered. Pass exactly one of: - csv_path: a .csv, .xlsx or .xls file inside a directory this server allows (it does the reading itself). Always prefer this -- it works for large files (up to the server's size limit, 200 MB by default) without the calling agent ever having to read or reproduce the file's contents itself. Excel files need the optional `glin-ml[excel]` install. - csv_content: the complete raw CSV text (header row + data rows). Only use this when there's no file on disk to point at, e.g. data pasted into the chat or assembled in memory. Do not use this for large files -- having an agent inline tens of thousands of rows as a tool-call argument is slow, burns enormous context, and risks truncation. Prefer writing the data to a file the server can see (csv_path) whenever that's possible instead. Excel only (csv_path ending in .xlsx/.xls): `sheet` picks the sheet by name or 1-based position (required when the workbook has several sheets). The header row is auto-detected, skipping title/banner rows and trailing footnote rows (mostly empty rows whose text starts with Source, Note, *, or is a long sentence; `warnings` names the dropped rows); if that is ambiguous the error says so. For multi-row or merged headers set `header_rows` to the 1-based sheet row numbers as shown in Excel: 4, "3-4" or [3, 4]; they are flattened into unique column names and data starts below them. The result reports the `sheet` and `header_rows` actually used. Either way, this tool does no cleanup of its own beyond what EBMTabularPreprocessor already does automatically (coercing dirty numeric formatting such as currency symbols, thousands separators and trailing % -- percentages sta…