Fix non-embedding param count (~92M) and drop unsourced token-count claim

README.md

@@ -107,8 +107,9 @@ count, temperature, and top-k.
 
 - Architecture: 12-layer DiT-style transformer, `d_model` 768, 12 heads, adaLN
   conditioning, untied embeddings, rotary positions. ≈169.6M total parameters
-  (≈130M non-embedding). Context length 1024.
-- Training: OpenWebText, 1M steps, ≈33B tokens (Sahoo et al., 2024).
+  (≈92M non-embedding; the input and output embeddings are untied, ≈77M together).
+  Context length 1024.
+- Training: OpenWebText, 1M steps (Sahoo et al., 2024).
 
 ## Requirements and environment