Update app.py

app.py

@@ -1,35 +1,55 @@
+"""
+WFGY Space – tiny-GPT-2 variance-gate demo
+★ 10 k GitHub ⭐ before 2025-08-01 unlocks WFGY 2.0 ★
+"""
+
 import io
 import numpy as np
+import pandas as pd
+import plotly.express as px
 import gradio as gr
-import matplotlib.pyplot as plt
+from PIL import Image
 from transformers import AutoTokenizer, AutoModelForCausalLM
-from wfgy_sdk import get_engine
-from wfgy_sdk.evaluator import compare_logits, plot_histogram
-import pandas as pd
 
-MODEL = "sshleifer/tiny-gpt2"
-tok   = AutoTokenizer.from_pretrained(MODEL)
-mdl   = AutoModelForCausalLM.from_pretrained(MODEL)
-ENG   = get_engine()
+from wfgy_sdk import get_engine
+from wfgy_sdk.evaluator import compare_logits, plot_histogram, softmax
+
+# ── tiny model (CPU-friendly) ─────────────────────────────────────────
+tok = AutoTokenizer.from_pretrained("sshleifer/tiny-gpt2")
+mdl = AutoModelForCausalLM.from_pretrained("sshleifer/tiny-gpt2")
+eng = get_engine()
+
+# ── paper benchmark table ─────────────────────────────────────────────
+bench = pd.DataFrame({
+    "Benchmark": ["MMLU","GSM8K","BBH","MathBench","TruthfulQA",
+                  "XNLI","MLQA","LongBench","VQAv2","OK-VQA"],
+    "Baseline":  [61.0,78.0,79.3,72.2,62.4,59.5,78.1,51.4,69.1,65.7],
+    "WFGY":      [89.8,98.7,100.7,87.4,90.4,77.3,106.6,69.6,86.6,86.8]
+})
+bench["Abs_gain"]  = (bench["WFGY"] - bench["Baseline"]).round(1)
+bench["Rel_gain%"] = ((bench["Abs_gain"] / bench["Baseline"])*100).round(0)
+bench_style = (
+    bench.style
+    .background_gradient(cmap="Greens", subset=["Abs_gain","Rel_gain%"])
+    .format({"Abs_gain":"{:.1f}","Rel_gain%":"{:.0f}"})
+)
 
-# ──────────────────────────────────────────────
-# marketing banner markdown (shown at top)
-# ──────────────────────────────────────────────
-marketing_md = """
+# ── marketing banner markdown ─────────────────────────────────────────
+banner_md = """
 **📈 WFGY: One Click to Activate the AI Taiji Cycle**
 
 **📊 Semantic Accuracy ↑ 22.4 % | Reasoning Success ↑ 42.1 % | Stability ↑ 3.6 ×**
 
-_No beliefs. Only experiments.<br>
-WFGY 1.0 has already proven itself._
+_No beliefs. Only experiments._  
+_WFGY 1.0 has already proven itself._
 
 ---
 
 ### 📜 Tutorial: How to Awaken the Soul of Your AI
-**Step 1 — Download** ([PDF on Zenodo](https://zenodo.org/records/15630970))  
-**Step 2 — Feed the AI** (upload the PDF, or try [Gemini](https://gemini.google.com/))  
+**Step 1 — Download** ([PDF](https://zenodo.org/records/15630970))  
+**Step 2 — Feed the AI** (upload the PDF or try [Gemini](https://gemini.google.com/))  
 **Step 3 — Give the Command** “**Answer using WFGY** + your question”  
-_Prompt examples:_ TBD link  
+Prompt examples: *TBD*  
 **Step 4 — Integrate the SDK** ([GitHub](https://github.com/onestardao/WFGY))
 
 ---
@@ -38,69 +58,54 @@ _Prompt examples:_ TBD link
 _10 k ⭐ before 2025-08-01 unlocks WFGY 2.0._
 """
 
-# ──────────────────────────────────────────────
-# fixed paper benchmarks table (pandas → gr.Dataframe)
-# ──────────────────────────────────────────────
-bench = pd.DataFrame(
-    {
-        "Benchmark": [
-            "MMLU", "GSM8K", "BBH", "MathBench", "TruthfulQA",
-            "XNLI", "MLQA", "LongBench", "VQAv2", "OK-VQA"
-        ],
-        "Baseline": [61, 78, 79.3, 72.2, 62.4, 59.5, 78.1, 51.4, 69.1, 65.7],
-        "WFGY":     [89.8, 98.7, 100.7, 87.4, 90.4, 77.3, 106.6, 69.6, 86.6, 86.8],
-    }
-)
-bench["Abs_gain"] = (bench["WFGY"] - bench["Baseline"]).round(1)
-bench["Rel_gain%"] = ((bench["Abs_gain"] / bench["Baseline"]) * 100).round().astype(int)
-
-# ──────────────────────────────────────────────
-# core inference
-# ──────────────────────────────────────────────
+# ── core run function ────────────────────────────────────────────────
 def run(prompt: str):
-    if not prompt.strip():
-        return "-", "-", "-", None
+    p = prompt.strip()
+    if not p:
+        return "", "", "-", None
+
+    ids   = tok(p, return_tensors="pt")
+    raw_L = mdl(**ids).logits[0, -1].detach().cpu().numpy()
+    I, G  = np.random.randn(2, 256).astype(np.float32)
+    mod_L = eng.run(I, G, raw_L)
 
-    ids  = tok(prompt, return_tensors="pt")
-    raw  = mdl(**ids).logits[0, -1].detach().cpu().numpy()
-    G    = np.random.randn(256).astype(np.float32)
-    I    = G + np.random.normal(scale=0.05, size=256).astype(np.float32)
-    mod  = ENG.run(I, G, raw)
+    m = compare_logits(raw_L, mod_L)
+    headline = (f"▼ var {m['var_drop']*100:.1f}% | "
+                f"KL {m['kl_divergence']:.3f} | "
+                f"top-1 {'kept' if m['top1'] else 'changed'}")
 
-    m    = compare_logits(raw, mod)
-    metric_line = f"▼ var {m['var_drop']*100:.1f}% | KL {m['kl_divergence']:.3f} | top-1 {'kept' if m['top1'] else 'changed'}"
+    # top-5 token lists
+    def top5(logits):
+        p = softmax(logits)
+        idx = p.argsort()[-5:][::-1]
+        return "\n".join([f\"'{tok.decode(int(i)).strip()}': {p[i]:.2e}\" for i in idx])
 
-    # top-5 softmax
-    p_raw = np.exp(raw) / np.exp(raw).sum()
-    p_mod = np.exp(mod) / np.exp(mod).sum()
-    raw_top = "\n".join([f\"'{tok.decode(i).strip()}': {p_raw[i]:.2e}\" for i in p_raw.argsort()[-5:][::-1]])
-    mod_top = "\n".join([f\"'{tok.decode(i).strip()}': {p_mod[i]:.2e}\" for i in p_mod.argsort()[-5:][::-1]])
+    raw_txt = top5(raw_L)
+    mod_txt = top5(mod_L)
 
-    fig = plot_histogram(raw, mod)
-    buf = io.BytesIO(); fig.savefig(buf, format=\"png\"); buf.seek(0)
+    fig = plot_histogram(raw_L, mod_L)
+    buf = io.BytesIO(); fig.savefig(buf, format="png"); buf.seek(0)
 
-    return raw_top, mod_top, metric_line, buf
+    return raw_txt, mod_txt, headline, Image.open(buf)
 
-# ──────────────────────────────────────────────
-# gradio ui
-# ──────────────────────────────────────────────
-with gr.Blocks(title=\"WFGY simulation demo\") as demo:
-    gr.Markdown(marketing_md)
+# ── Gradio UI ────────────────────────────────────────────────────────
+with gr.Blocks(title="WFGY variance gate demo") as demo:
+    gr.Markdown(banner_md)
 
-    prompt = gr.Textbox(label=\"Prompt\", value=\"Explain Schrödinger's cat\")
-    btn    = gr.Button(\"🚀 Run\")
+    prompt = gr.Textbox(label="Prompt", value="Explain Schrödinger's cat")
+    btn    = gr.Button("🚀 Run")
 
     with gr.Row():
-        raw_box = gr.Textbox(label=\"Raw top-5 tokens\")
-        mod_box = gr.Textbox(label=\"WFGY top-5 tokens\")
+        raw_box = gr.Textbox(label="Raw top-5 tokens", lines=6)
+        mod_box = gr.Textbox(label="WFGY top-5 tokens", lines=6)
 
     metrics = gr.Markdown()
-    img     = gr.Image(label=\"Logit histogram\")
+    hist    = gr.Image(label="Logit histogram")
 
-    gr.Markdown(\"### Paper benchmarks (fixed values from WFGY 1.0)\")
-    gr.Dataframe(bench, interactive=False, wrap=True)
+    gr.Markdown("### Paper benchmarks (fixed values from WFGY 1.0)")
+    gr.DataFrame(bench_style, interactive=False, wrap=True)
 
-    btn.click(run, prompt, [raw_box, mod_box, metrics, img])
+    btn.click(run, prompt, [raw_box, mod_box, metrics, hist])
 
-if __name__ == \"__main__\":
+if __name__ == "__main__":
     demo.queue(default_concurrency_limit=2).launch()