Add reward_train_max_length: 512 so the calibration regime ships with the weights

config.json

@@ -36,10 +36,11 @@
   "position_biased_input": false,
   "position_buckets": 256,
   "relative_attention": true,
-  "reward_temperature": 1.166796088218689,
   "share_att_key": true,
   "tie_word_embeddings": true,
   "transformers_version": "5.12.1",
   "type_vocab_size": 0,
-  "vocab_size": 128100
-}
+  "vocab_size": 128100,
+  "reward_temperature": 1.166796088218689,
+  "reward_train_max_length": 512
+}