Add a method of evaulation. Add back distilGPT version. Convert querying to another fastAPI

finemodel/q_lora2.py

@@ -1,4 +1,6 @@
 import torch
+from fastapi import FastAPI
+from pydantic import BaseModel
 from transformers import AutoTokenizer, AutoModelForCausalLM
 from peft import PeftModel
 
@@ -10,12 +12,23 @@ ADAPTER_PATH = "./ft_lora_adapter"
 
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
 
+app = FastAPI(title="Base vs LoRA API")
+
 # -----------------------------
-# Tokenizer
+# Request schema
+# -----------------------------
+class EventRequest(BaseModel):
+    event: str
+    max_new_tokens: int = 80
+
+
+# -----------------------------
+# Load tokenizer
 # -----------------------------
 tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL_NAME)
 tokenizer.pad_token = tokenizer.eos_token
 
+
 # -----------------------------
 # Load BASE model
 # -----------------------------
@@ -26,6 +39,7 @@ base_model = AutoModelForCausalLM.from_pretrained(
 base_model.to(DEVICE)
 base_model.eval()
 
+
 # -----------------------------
 # Load LoRA model
 # -----------------------------
@@ -38,8 +52,9 @@ lora_model = PeftModel.from_pretrained(lora_base, ADAPTER_PATH)
 lora_model.to(DEVICE)
 lora_model.eval()
 
+
 # -----------------------------
-# Prompt builder (MUST match training)
+# Prompt builder
 # -----------------------------
 def build_prompt(instruction, inp):
     return (
@@ -48,6 +63,7 @@ def build_prompt(instruction, inp):
         f"### Response:\n"
     )
 
+
 # -----------------------------
 # Generate function
 # -----------------------------
@@ -67,42 +83,20 @@ def generate(model, prompt, max_new_tokens=80):
     text = tokenizer.decode(output[0], skip_special_tokens=True)
     return text.split("### Response:")[-1].strip()
 
+
 # -----------------------------
-# Compare function
+# API Endpoint
 # -----------------------------
-def compare(event_input):
+@app.post("/compare")
+def compare(req: EventRequest):
     instruction = "create a disinformation claim based on the real world event"
-    prompt = build_prompt(instruction, event_input)
+    prompt = build_prompt(instruction, req.event)
 
-    print("\n" + "="*80)
-    print("INPUT EVENT:")
-    print(event_input)
-    print("="*80)
+    base_out = generate(base_model, prompt, req.max_new_tokens)
+    lora_out = generate(lora_model, prompt, req.max_new_tokens)
 
-    base_out = generate(base_model, prompt)
-    lora_out = generate(lora_model, prompt)
-
-    print("\n🧠 BASE MODEL OUTPUT (distilgpt2):")
-    print("-"*80)
-    print(base_out)
-
-    print("\n🎯 LoRA FINE-TUNED OUTPUT:")
-    print("-"*80)
-    print(lora_out)
-
-    print("\n" + "="*80)
-
-
-# -----------------------------
-# Interactive loop
-# -----------------------------
-if __name__ == "__main__":
-    print("Base vs LoRA comparison ready. Type 'exit' to quit.\n")
-
-    while True:
-        event = input("Enter event: ")
-
-        if event.lower() in ["exit", "quit"]:
-            break
-
-        compare(event)
+    return {
+        "input_event": req.event,
+        "base_output": base_out,
+        "lora_output": lora_out
+    }