Upload README.md with huggingface_hub

README.md

@@ -65,13 +65,15 @@ request. Two automatic filters applied to every generated surface: rejection of
 paraphrases too close to their source question (token-set Jaccard >= 0.75), and
 near-duplicate removal across questions.
 
-**No-leakage training methodology.** Our fine-tuning experiments do not train on this
-published pool. Training corpora are built separately from train-split questions only,
-and every candidate surface — original or paraphrase — is screened against the private
-held-out evaluation set with a token-set Jaccard firewall at 0.7: any surface at or
-above the threshold is dropped. An independent verification step re-checks the final
-training file after every build; the maximum retained similarity to date is 0.667,
-below threshold. The held-out set itself is never trained on and is not published.
+**No-leakage training methodology.** This published pool covers every source
+question, including paraphrases of questions related to our held-out evaluation set.
+For our own fine-tuning we therefore use a filtered subset of the same generation
+output: only paraphrases of train-split questions, after screening every candidate
+surface — original or paraphrase — against the private held-out evaluation set with a
+token-set Jaccard firewall at 0.7 (any surface at or above the threshold is dropped).
+An independent verification step re-checks the final training file after every build;
+no training surface is a held-out question, a paraphrase of one, or near-duplicate of
+one (maximum retained similarity to date: 0.667, below threshold).
 
 **Generalization evidence.** Fine-tuning on the decontaminated training subset
 improved accuracy on held-out questions over the base model, indicating the