Upload train_unsloth.py with huggingface_hub

train_unsloth.py

@@ -12,9 +12,11 @@
 # ///
 
 import os
+
+# Import unsloth FIRST before anything else
+from unsloth import FastLanguageModel
 from datasets import load_dataset
 from trl import SFTTrainer, SFTConfig
-from unsloth import FastLanguageModel
 from transformers import AutoTokenizer
 import trackio
 
@@ -41,10 +43,13 @@ model, _ = FastLanguageModel.from_pretrained(
     trust_remote_code=True,
 )
 
-# Load tokenizer directly and fix eos_token
+# Load tokenizer directly
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL, trust_remote_code=True)
+# Explicitly set special tokens to avoid TokenizersBackend defaults
 tokenizer.eos_token = "<|im_end|>"
 tokenizer.pad_token = tokenizer.eos_token
+tokenizer.pad_token_id = tokenizer.convert_tokens_to_ids(tokenizer.pad_token)
+tokenizer.eos_token_id = tokenizer.convert_tokens_to_ids(tokenizer.eos_token)
 
 model = FastLanguageModel.get_peft_model(
     model,
@@ -88,6 +93,8 @@ trainer = SFTTrainer(
         max_length=1024,
         bf16=True,
         dataset_text_field="text",
+        packing=False,
+        eos_token="<|im_end|>",
     ),
 )