Sync from GitHub: 0c5737d7108460e1c2b09e575eccf57cc50766be

grammars/concepts.gbnf

@@ -28,5 +28,14 @@
 
 root    ::= item ("," ws item){0,7}
 item    ::= word (ws word){0,3}
-word    ::= [a-zA-Z] [a-zA-Z0-9-]{2,19}
+# Word: first char is any letter (allows TitleCase proper nouns
+# like "Calvin" or "RSA"); body is lowercase + digits + hyphen only.
+# Mid-word capitals are forbidden. This forces Falcon3's BPE
+# tokenizer to emit space-prefixed continuation tokens (" Sunlight")
+# rather than jamming tokens together as one camelCase "word"
+# ("thatGreenhouseCarbon"). Per arXiv 2502.14969, space-prefixed
+# tokens have better-trained embeddings (5-10% gain), which is
+# particularly important for smaller/quantized models like
+# Falcon3-10B-1.58bit.
+word    ::= [a-zA-Z] [a-z0-9-]{2,19}
 ws      ::= " "