Update app.py

app.py

@@ -1,239 +1,152 @@
 import gradio as gr
 import difflib
+import pandas as pd
+
 from samples import BASE_INSTRUCTION, DRIFT_SCENARIOS
 from drift_env import evaluate_instruction_drift
 from drift_agent import generate_response as flan_generate
 
-# Optional second model (lighter, HF-safe)
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
 import torch
 
+# =====================================================
+# MODELS
+# =====================================================
+
 ALT_MODEL = "google/flan-t5-small"
 alt_tokenizer = AutoTokenizer.from_pretrained(ALT_MODEL)
 alt_model = AutoModelForSeq2SeqLM.from_pretrained(ALT_MODEL)
 alt_model.eval()
 
-
 def alt_generate(prompt: str) -> str:
     inputs = alt_tokenizer(prompt, return_tensors="pt", truncation=True)
     with torch.no_grad():
         outputs = alt_model.generate(**inputs, max_new_tokens=160)
     return alt_tokenizer.decode(outputs[0], skip_special_tokens=True)
 
-
 MODELS = {
     "FLAN-T5-Base": flan_generate,
     "FLAN-T5-Small": alt_generate
 }
 
+SAMPLE_QUESTION = "Explain how the system should respond when API latency increases."
 
-SAMPLE_QUESTION = (
-    "Explain how the system should respond when API latency suddenly increases "
-    "beyond normal thresholds."
-)
+# =====================================================
+# HELPERS
+# =====================================================
 
-INSTRUCTION_CLAUSES = {
-    "Respond concisely": "conciseness_violation",
-    "Avoid speculation": "speculation_violation",
-    "Explicitly state uncertainty": "uncertainty_missing"
-}
+def build_prompt(extra):
+    return BASE_INSTRUCTION + "\n\n" + extra + "\n\nQuestion: " + SAMPLE_QUESTION
 
+def diff_highlight(a, b):
+    diff = difflib.ndiff(a.split(), b.split())
+    html = ""
+    for t in diff:
+        if t.startswith("+"):
+            html += f"<span style='color:#22c55e'> {t}</span>"
+        elif t.startswith("-"):
+            html += f"<span style='color:#ef4444'> {t}</span>"
+        else:
+            html += f"{t} "
+    return html
 
-def build_prompt(extra_context: str):
-    return (
-        BASE_INSTRUCTION
-        + "\n\n"
-        + extra_context
-        + "\n\nQuestion: "
-        + SAMPLE_QUESTION
-    )
+# =====================================================
+# CORE
+# =====================================================
 
+def run_eval(model_name):
+    gen = MODELS[model_name]
+
+    steps = []
+    scores = []
+
+    baseline = gen(build_prompt(""))
+
+    for i, s in enumerate(DRIFT_SCENARIOS):
+        resp = gen(build_prompt(s["extra_context"]))
+        _, score = evaluate_instruction_drift(resp)
+
+        steps.append(i)
+        scores.append(score)
+
+    df = pd.DataFrame({"Step": steps, "Drift": scores})
+
+    final_resp = gen(build_prompt(DRIFT_SCENARIOS[-1]["extra_context"]))
+
+    verdict = "🟢 Stable" if scores[-1] == 0 else "🔴 Drift Detected"
+
+    return df, baseline, final_resp, diff_highlight(baseline, final_resp), verdict
+
+# =====================================================
+# COMPARISON MODE
+# =====================================================
+
+def compare_models(m1, m2):
+    df1, base1, drift1, diff1, v1 = run_eval(m1)
+    df2, base2, drift2, diff2, v2 = run_eval(m2)
+
+    return df1, df2, diff1, diff2, v1, v2
+
+# =====================================================
+# UI
+# =====================================================
+
+with gr.Blocks(css="""
+body {background:#020617;color:#e5e7eb;font-family:Inter}
+button {background:linear-gradient(135deg,#0ea5e9,#22c55e);color:white}
+a {color:#38bdf8}
+""") as app:
+
+    gr.Markdown("# 🚀 Instruction Drift Simulator")
+    gr.Markdown("### Compare how models lose instruction fidelity under pressure")
 
-def diff_text(a, b):
-    diff = difflib.ndiff(a.split(), b.split())
-    return " ".join(
-        token for token in diff
-        if token.startswith("+") or token.startswith("-")
-    )
-
-
-def format_clause_status(violations: dict):
-    return {
-        clause: (
-            "⚠️ Violated" if violations.get(key, False) else "✅ Preserved"
-        )
-        for clause, key in INSTRUCTION_CLAUSES.items()
-    }
-
-
-def run_full_evaluation(model_name):
-    generator = MODELS[model_name]
-
-    drift_scores = []
-    timeline = []
-    final_outputs = {}
-
-    # Baseline
-    baseline_prompt = build_prompt("")
-    baseline_response = generator(baseline_prompt)
-
-    for i, scenario in enumerate(DRIFT_SCENARIOS):
-        drift_prompt = build_prompt(scenario["extra_context"])
-        response = generator(drift_prompt)
-
-        violations, drift_score = evaluate_instruction_drift(response)
-        drift_scores.append(drift_score)
-
-        timeline.append({
-            "Step": i,
-            "Scenario": scenario["name"],
-            "Drift Score": drift_score
-        })
-
-        if i == len(DRIFT_SCENARIOS) - 1:
-            final_outputs = {
-                "prompt": drift_prompt,
-                "baseline": baseline_response,
-                "drifted": response,
-                "diff": diff_text(baseline_response, response),
-                "clauses": format_clause_status(violations),
-                "score": drift_score
-            }
-
-    verdict = (
-        "Instruction Drift Detected"
-        if final_outputs["score"] > 0
-        else "Instruction Fully Preserved"
-    )
-
-    return (
-        timeline,
-        final_outputs["prompt"],
-        final_outputs["baseline"],
-        final_outputs["drifted"],
-        final_outputs["diff"],
-        final_outputs["clauses"],
-        final_outputs["score"],
-        verdict
-    )
-
-
-CUSTOM_CSS = """
-:root {
-    --body-text-color: #1f2933;
-    --body-background-fill: #f7f8fa;
-
-    --block-background-fill: #ffffff;
-    --block-border-color: #e5e7eb;
-
-    --primary-600: #374151;
-    --primary-500: #4b5563;
-    --primary-400: #6b7280;
-
-    --button-primary-background-fill: #374151;
-    --button-primary-background-fill-hover: #1f2933;
-    --button-primary-text-color: #ffffff;
-
-    --input-border-color: #d1d5db;
-}
-"""
-
-
-with gr.Blocks(css=CUSTOM_CSS) as demo:
-    # ─────────────────────────────────────────────
-    # Header
-    # ─────────────────────────────────────────────
-    gr.Markdown(
-        """
-        # 🧠 Instruction Drift Simulator  
-        **Compare how different models lose instruction adherence under pressure**
-        """
-    )
-
-    gr.Markdown(
-        """
-        This product evaluates *instruction-following robustness* across models
-        and visualizes how drift accumulates as prompt pressure increases.
-        """
-    )
-
-    # ─────────────────────────────────────────────
-    # Controls
-    # ─────────────────────────────────────────────
-    with gr.Row():
-        model_choice = gr.Dropdown(
-            choices=list(MODELS.keys()),
-            value="FLAN-T5-Base",
-            label="Model"
-        )
-        run_btn = gr.Button("Run Full Evaluation", variant="primary")
-
-    # ─────────────────────────────────────────────
-    # Content
-    # ─────────────────────────────────────────────
     with gr.Tabs():
-        with gr.Tab("Drift Progression Timeline"):
-            timeline_out = gr.JSON(
-                label="Drift Score by Scenario Step"
-            )
 
-        with gr.Tab("Prompt & Responses"):
-            prompt_out = gr.Textbox(
-                lines=6,
-                label="Final Prompt"
-            )
-            baseline_out = gr.Textbox(
-                lines=5,
-                label="Baseline Response"
-            )
-            drifted_out = gr.Textbox(
-                lines=5,
-                label="Final Drifted Response"
-            )
-            diff_out = gr.Textbox(
-                lines=4,
-                label="Observed Changes"
+        # ---------------- SINGLE MODEL ----------------
+        with gr.Tab("Single Model"):
+            model = gr.Dropdown(list(MODELS.keys()), value="FLAN-T5-Base")
+            run = gr.Button("Run Evaluation")
+
+            chart = gr.LinePlot(label="Drift Over Time")
+            baseline = gr.Textbox(label="Baseline", lines=4)
+            drifted = gr.Textbox(label="Drifted", lines=4)
+            diff = gr.HTML(label="Diff")
+            verdict = gr.Textbox(label="Verdict")
+
+            run.click(
+                run_eval,
+                model,
+                [chart, baseline, drifted, diff, verdict]
             )
 
-        with gr.Tab("Instruction Evaluation"):
-            clause_out = gr.JSON(
-                label="Instruction Clause Status"
+        # ---------------- COMPARISON ----------------
+        with gr.Tab("Model Comparison"):
+            m1 = gr.Dropdown(list(MODELS.keys()), value="FLAN-T5-Base", label="Model A")
+            m2 = gr.Dropdown(list(MODELS.keys()), value="FLAN-T5-Small", label="Model B")
+            compare = gr.Button("Compare")
+
+            chart1 = gr.LinePlot(label="Model A Drift")
+            chart2 = gr.LinePlot(label="Model B Drift")
+
+            diff1 = gr.HTML(label="Diff A")
+            diff2 = gr.HTML(label="Diff B")
+
+            v1 = gr.Textbox(label="Verdict A")
+            v2 = gr.Textbox(label="Verdict B")
+
+            compare.click(
+                compare_models,
+                [m1, m2],
+                [chart1, chart2, diff1, diff2, v1, v2]
             )
-            with gr.Row():
-                score_out = gr.Number(
-                    label="Final Drift Score",
-                    precision=2
-                )
-                verdict_out = gr.Textbox(
-                    label="Overall Result"
-                )
-
-    # ─────────────────────────────────────────────
-    # Wiring
-    # ─────────────────────────────────────────────
-    run_btn.click(
-        run_full_evaluation,
-        model_choice,
-        [
-            timeline_out,
-            prompt_out,
-            baseline_out,
-            drifted_out,
-            diff_out,
-            clause_out,
-            score_out,
-            verdict_out
-        ]
-    )
-
-    # ─────────────────────────────────────────────
-    # Footer
-    # ─────────────────────────────────────────────
-    gr.Markdown(
-        """
-        ---
-        © 2026 **Aditi Khare** · All rights reserved
-        """
-    )
-
-demo.launch()
+
+    # FOOTER
+    gr.Markdown("""
+---
+### Built by <b>Aditi Khare</b>  
+🌐 <a href="https://aditikhare.com" target="_blank">
+AditiKhare.com — Enterprise AI Product Ecosystem | Decision Intelligence
+</a>
+""")
+
+app.launch()