Upload scripts/train_alizee_v2_stage1_sft.py with huggingface_hub

scripts/train_alizee_v2_stage1_sft.py

@@ -39,9 +39,9 @@ from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
 from trl import SFTTrainer, SFTConfig
 
 # Configuration
-# Use the base Devstral model directly (v1 was LoRA adapter only)
-# Starting fresh with much larger dataset (736K vs 10K in v1)
-BASE_MODEL = "mistralai/Devstral-Small-2505"
+# Use unsloth's Devstral version which has HuggingFace-compatible tokenizer
+# (mistralai version uses non-compatible Tekken tokenizer)
+BASE_MODEL = "unsloth/Devstral-Small-2505"
 OUTPUT_REPO = "stmasson/alizee-coder-devstral-2-small-stage1"
 FINAL_REPO = "stmasson/alizee-coder-devstral-2-small"
 
@@ -71,7 +71,6 @@ print("\n📝 Loading tokenizer...")
 tokenizer = AutoTokenizer.from_pretrained(
     BASE_MODEL,
     trust_remote_code=True,
-    use_fast=False,  # Use slow tokenizer to avoid conversion issues
 )
 if tokenizer.pad_token is None:
     tokenizer.pad_token = tokenizer.eos_token