Update app.py

app.py

@@ -4,11 +4,13 @@ from transformers import AutoModelForCausalLM, AutoTokenizer
 from peft import PeftModel
 
 # 1. Configuration
+# We point to the official unsloth version of the base model for the tokenizer
 base_model_name = "unsloth/Llama-3.2-3B-bnb-4bit"
 adapter_path = "." 
 
-# 2. Load Tokenizer with fix for Llama-3 (no Unsloth)
-tokenizer = AutoTokenizer.from_pretrained(adapter_path, use_fast=False)
+# 2. Load Tokenizer directly from the base model source
+# This fixes the "backend tokenizer" error by getting the files from Hugging Face directly
+tokenizer = AutoTokenizer.from_pretrained(base_model_name)
 
 # 3. Load Model with memory-saving settings for CPU
 model = AutoModelForCausalLM.from_pretrained(
@@ -18,10 +20,10 @@ model = AutoModelForCausalLM.from_pretrained(
     low_cpu_mem_usage=True
 )
 
-# 4. Apply your adapters
+# 4. Apply your fine-tuned adapters
 model = PeftModel.from_pretrained(model, adapter_path)
 
-# 5. Logic
+# 5. The Reasoning Logic
 def legal_summarizer(legal_text):
     prompt = f"Analyze the following legal text and provide a grounded summary.\n\nInput:\n{legal_text}\n\nResponse:\n"
     inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
@@ -35,7 +37,7 @@ def legal_summarizer(legal_text):
         return decoded.split("Response:")[-1].strip()
     return decoded.strip()
 
-# 6. UI
+# 6. Setup the UI
 demo = gr.Interface(
     fn=legal_summarizer, 
     inputs=gr.Textbox(lines=10, label="Paste Legal Clause"),