model card

README.md

@@ -0,0 +1,32 @@
+---
+license: apache-2.0
+tags: [ai-detection, logistic-regression]
+extra_gated_prompt: "Access is granted individually. Please say who you are and what you intend to use the weights for."
+---
+
+# ideadet-logreg-1m-outline
+
+Idea-level AI-text detector: a logistic-regression head over frozen
+`text-embedding-3-large` vectors (3072 dimensions). It answers **whose ideas a
+document contains**, not who typed the sentences.
+
+- **Training corpus:** 1m, 842,301 training rows
+- **Reads at inference:** a de-leaked (paraphrased) role-labelled outline, embedded whole
+- **Hyper-parameters:** C=1.0, max_iter=3000
+
+## Contents
+
+`lr_full_v1m.npz` holds `coef` ((1, 3072)), `intercept`, and the fitting metadata.
+AI is the positive class, so the reported score is `predict_proba(x)[:, 0]` = P(human), and
+the detector fires when that falls below a calibrated threshold. Thresholds are quantiles
+over held-out human documents and are **not** included here: a cut from one model or input
+form is meaningless against another's scores.
+
+## Provenance
+
+This model was refit from the stored embeddings, because the original runs saved only their
+predictions. The refit was checked against those published test predictions and reproduces
+them to **max |difference| = 2.06e-13**, so it is the same
+model that produced the reported numbers rather than an approximation.
+
+Fuller documentation and evaluation results to follow.