Upload train_question_generator.py with huggingface_hub

train_question_generator.py

@@ -1,5 +1,5 @@
 # /// script
-# dependencies = ["trl>=0.17.0", "peft>=0.15.0", "datasets", "transformers", "accelerate", "bitsandbytes"]
+# dependencies = ["trl>=0.12.0", "peft>=0.13.0", "datasets", "transformers>=4.45.0", "accelerate", "bitsandbytes", "huggingface_hub"]
 # ///
 
 import os
@@ -9,6 +9,13 @@ from trl import SFTTrainer, SFTConfig
 from transformers import AutoTokenizer, AutoModelForCausalLM, BitsAndBytesConfig
 import torch
 
+# Authenticate
+from huggingface_hub import login
+hf_token = os.environ.get("HF_TOKEN")
+if hf_token:
+    login(token=hf_token)
+    print("Authenticated with HuggingFace")
+
 print("Loading dataset...")
 dataset = load_dataset("KevinKeller/cognitive-question-generator-v1")
 train_dataset = dataset["train"]
@@ -18,7 +25,7 @@ print(f"Train samples: {len(train_dataset)}")
 if eval_dataset:
     print(f"Eval samples: {len(eval_dataset)}")
 
-# Using Qwen2.5-7B for question generation (good reasoning capabilities)
+# Using Qwen2.5-7B for question generation
 print("Loading model: Qwen/Qwen2.5-7B-Instruct...")
 model_id = "Qwen/Qwen2.5-7B-Instruct"
 
@@ -30,7 +37,8 @@ bnb_config = BitsAndBytesConfig(
 )
 
 tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
-tokenizer.pad_token = tokenizer.eos_token
+if tokenizer.pad_token is None:
+    tokenizer.pad_token = tokenizer.eos_token
 
 model = AutoModelForCausalLM.from_pretrained(
     model_id,
@@ -75,7 +83,7 @@ trainer = SFTTrainer(
     train_dataset=train_dataset,
     eval_dataset=eval_dataset,
     peft_config=peft_config,
-    processing_class=tokenizer,
+    tokenizer=tokenizer,
     args=training_args,
 )