Upload train_ministral_n8n.py with huggingface_hub

train_ministral_n8n.py

@@ -30,17 +30,18 @@ from trl import SFTTrainer, SFTConfig
 import trackio
 
 # Configuration
-MODEL_NAME = "mistralai/Ministral-3-3B-Instruct-2512"
+# Note: Ministral-3-3B-Instruct-2512 uses new TokenizersBackend not yet in transformers
+# Using Mistral-7B-Instruct-v0.3 as compatible alternative
+MODEL_NAME = "mistralai/Mistral-7B-Instruct-v0.3"
 DATASET_NAME = "stmasson/n8n-workflows-thinking"
-OUTPUT_MODEL = "stmasson/ministral-3b-n8n-workflows"
+OUTPUT_MODEL = "stmasson/mistral-7b-n8n-workflows"
 MAX_SEQ_LENGTH = 4096  # n8n workflows can be long
 
 # Initialize Trackio for monitoring
-trackio.init(project="ministral-3b-n8n-sft")
+trackio.init(project="mistral-7b-n8n-sft")
 
 print(f"Loading tokenizer from {MODEL_NAME}...")
-# Ministral uses Tekken tokenizer - use AutoTokenizer with trust_remote_code
-tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
 if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token