driftless 0.1.0__py3-none-any.whl

driftless/__init__.py

@@ -0,0 +1,3 @@
+"""driftless: Dependabot for LLM models."""
+
+__version__ = "0.1.0"

driftless/calibrate.py

@@ -0,0 +1,33 @@
+"""Baseline-derived threshold suggestions (the educational half of P2.2).
+
+A first-time user can't guess reasonable ``min_f1`` / ``max_schema_error_rate``
+values. ``suggest_thresholds`` turns measured baseline metrics into a starting
+``thresholds:`` block (achieved metric minus a safety margin), which the user can
+accept or edit. Pure and reused by the ``calibrate`` CLI command.
+"""
+
+from __future__ import annotations
+
+from .evaluation import Metrics
+
+DEFAULT_MARGIN = 0.03
+
+
+def suggest_thresholds(metrics: Metrics, *, margin: float = DEFAULT_MARGIN) -> dict:
+    """Suggested absolute thresholds grounded in measured baseline metrics.
+
+    Only emits a key when the underlying metric was actually measured, so we
+    never invent a bar for something we couldn't evaluate.
+    """
+    out: dict[str, float] = {}
+    if metrics.f1 is not None:
+        out["min_f1"] = round(max(0.0, metrics.f1 - margin), 3)
+    if metrics.score is not None:
+        out["min_score"] = round(max(0.0, metrics.score - margin), 3)
+    if metrics.precision is not None:
+        out["min_precision"] = round(max(0.0, metrics.precision - margin), 3)
+    if metrics.recall is not None:
+        out["min_recall"] = round(max(0.0, metrics.recall - margin), 3)
+    if metrics.schema_error_rate is not None:
+        out["max_schema_error_rate"] = round(min(1.0, metrics.schema_error_rate + margin), 3)
+    return out