Upload train_ministral_n8n.py with huggingface_hub

train_ministral_n8n.py

@@ -37,9 +37,9 @@ MAX_SEQ_LENGTH = 4096  # n8n workflows can be long
 trackio.init(project="ministral-3b-n8n-sft")
 
 print(f"Loading tokenizer from {MODEL_NAME}...")
-# Ministral uses a new tokenizer backend - load with use_fast=True
-from transformers import MistralTokenizerFast
-tokenizer = MistralTokenizerFast.from_pretrained(MODEL_NAME)
+# Ministral uses tekken tokenizer - use PreTrainedTokenizerFast
+from transformers import PreTrainedTokenizerFast
+tokenizer = PreTrainedTokenizerFast.from_pretrained(MODEL_NAME)
 if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token