Fix: base tokenizer, set bos/eos/pad, decoder_start_token_id=bos, forced_bos_token_id=mn_MN

config.json

@@ -16,7 +16,7 @@
   "decoder_ffn_dim": 4096,
   "decoder_layerdrop": 0.0,
   "decoder_layers": 12,
-  "decoder_start_token_id": 250037,
+  "decoder_start_token_id": 0,
   "dropout": 0.1,
   "dtype": "float32",
   "early_stopping": null,