Lead the model card with the issue-localization result

README.md

@@ -39,6 +39,11 @@ model-index:
 
 # CodeIntel-150M
 
+> **About 55% better than any other open ~150M embedding model at finding the code a bug report is about.**
+> On 3,804 held-out GitHub issues that name no code, CodeIntel scores nDCG@10 **0.268 / 0.274** (Python /
+> TypeScript) against **0.173 / 0.178** for the next-best of seven open models its size. It also leads on issues
+> that do name code (+27%) and on test -> implementation search (+23%).
+
 A 149M-parameter embedding model for **finding the code a bug report or feature request is about**. Given an issue
 written by a user, it retrieves the functions and classes in a Python or TypeScript repository that a fix will touch.
 It also handles code-to-code navigation (test -> implementation, call site -> definition) and code search
@@ -57,7 +62,7 @@ Apache-2.0). It runs locally, including as a 325 MB INT8 ONNX model, and needs n
 ## Results
 
 All numbers come from held-out test data that no training or model-selection step used. Each model was scored once,
-with the same protocol. Comparisons are with open models of similar size, all zero-shot.
+with the same protocol. The comparison models are open models of similar size, evaluated zero-shot.
 
 **Issue -> code localization.** The test set has 3,804 GitHub issues from 35 Python and TypeScript repositories, none
 of which were used in training. A hit means retrieving a function or class that the issue's fix edited, among all