Update app.py

app.py

@@ -1,58 +1,50 @@
-import gradio as gr
-from transformers import pipeline
+import os, json, gradio as gr
+from huggingface_hub import InferenceClient
 
-# Load a Hugging Face instruction-tuned model
-generator = pipeline("text-generation", model="tiiuae/falcon-7b-instruct", max_new_tokens=256)
+MODEL_ID = "tiiuae/falcon-7b-instruct"  # keep the big model
+client = InferenceClient(model=MODEL_ID, token=os.getenv("HF_TOKEN"))
 
-def evaluate_reu_problem(problem_statement):
-    prompt = f"""
-You are an evaluator of undergraduate research problem statements.
-Return the following checklist JSON with clear Yes/No and 0-3 scores.
-
-Checklist:
-{{
-  "G1": "", "G2": "", "G3": "", "G4": "", "G5": "",
-  "S1": 0, "S2": 0, "S3": 0, "S4": 0, "S5": 0,
-  "comments": ""
-}}
-
-Problem statement: \"\"\"{problem_statement}\"\"\"
-
-Checklist:
+SYSTEM_PROMPT = """You are an evaluator of undergraduate research problem statements.
+Return only a JSON object with the following keys:
+G1,G2,G3,G4,G5 (Yes/No) and S1–S5 (0-3) and comments.
 """
-    result = generator(prompt)[0]["generated_text"]
 
-    # Extract only JSON (if possible)
-    json_start = result.find("{")
-    checklist_json = result[json_start:].strip() if json_start != -1 else "Unable to parse JSON."
+def assess(statement):
+    prompt = f"{SYSTEM_PROMPT}\nProblem statement:\n\"\"\"{statement}\"\"\"\nChecklist:"
+    response = client.text_generation(
+        prompt,
+        max_new_tokens=256,
+        temperature=0.1,
+        top_p=0.9,
+        stream=False,
+    )
+
+    # Get first JSON block
+    start = response.find("{")
+    end   = response.rfind("}") + 1
+    json_str = response[start:end] if start != -1 else "{}"
 
-    # Evaluate decision logic (simple heuristic)
     try:
-        checklist = eval(checklist_json)
-        if any(checklist.get(f"G{i}", "No") == "No" for i in range(1, 6)):
-            decision = "❌ Does NOT meet requirements. Please revise the problem statement."
+        checklist = json.loads(json_str)
+        gate_fail = any(checklist.get(f"G{i}") == "No" for i in range(1,6))
+        score     = sum(int(checklist.get(f"S{i}",0)) for i in range(1,6))
+        if gate_fail:
+            verdict = "❌ Mandatory gatekeeper criterion failed."
+        elif score >= 11:
+            verdict = "✅ Meets requirements (Exceptional/Proficient)."
+        elif score >= 8:
+            verdict = "⚠️ Adequate but needs refinement."
         else:
-            score = sum(checklist.get(f"S{i}", 0) for i in range(1, 6))
-            if score >= 11:
-                decision = "✅ Meets requirements (Exceptional)"
-            elif score >= 8:
-                decision = "✅ Meets requirements (Proficient)"
-            else:
-                decision = "⚠️ Needs Improvement"
-    except:
-        decision = "⚠️ Could not evaluate due to JSON parsing error."
-
-    return checklist_json, decision
+            verdict = "❌ Needs major improvement."
+    except Exception as e:
+        verdict = f"⚠️ JSON parse error: {e}"
+    return json_str, verdict
 
 iface = gr.Interface(
-    fn=evaluate_reu_problem,
-    inputs=gr.Textbox(label="Enter REU Problem Statement", lines=7),
-    outputs=[
-        gr.Textbox(label="Checklist (JSON Output)", lines=15),
-        gr.Textbox(label="Final Assessment")
-    ],
-    title="🧠 REU Research Problem Evaluator",
-    description="Paste a research problem statement. The evaluator will fill the feasibility checklist and recommend action."
+    fn=assess,
+    inputs=gr.Textbox(lines=8, label="Paste REU Problem Statement"),
+    outputs=[gr.Code(label="Checklist JSON"), gr.Textbox(label="Verdict")],
+    title="🧠 REU Problem Statement Evaluator",
+    description="Uses the Hugging Face Inference API so no large model is loaded in the Space itself."
 )
-
 iface.launch()