Update app.py

app.py

@@ -1,12 +1,3 @@
-"""
-WFGY HuggingFace Space — deluxe marketing demo
-----------------------------------------------
-* Show before/after text
-* Display variance drop, KL, top-1 shift
-* Overlay histogram
-* Rich Markdown explaining every metric, PDF trick, star goal, secret papers
-"""
-
 import base64, io, numpy as np, gradio as gr, wfgy_sdk as w
 from wfgy_sdk.evaluator import compare_logits
 from wfgy_sdk.visual import plot_histogram
@@ -14,119 +5,102 @@ from wfgy_sdk.visual import plot_histogram
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed
 
-MODEL = "sshleifer/tiny-gpt2"          # fast CPU model
+MODEL = "sshleifer/tiny-gpt2"
 tokenizer = AutoTokenizer.from_pretrained(MODEL)
-model     = AutoModelForCausalLM.from_pretrained(MODEL)
+model = AutoModelForCausalLM.from_pretrained(MODEL)
 set_seed(42)
+ENGINE = w.get_engine()
 
-ENGINE = w.get_engine()                # singleton
 
-
-# ------------------------------------------------------------
-# helper: run WFGY or bypass, return text + metrics + img
-# ------------------------------------------------------------
-def wfgy_demo(prompt: str, enable_wfgy: bool):
+def wfgy_pipeline(prompt: str, enable_wfgy: bool):
     if not prompt.strip():
-        return "", "", "", ""
-
-    # ----- raw logits -----
-    ids = tokenizer(prompt, return_tensors="pt").input_ids
-    with torch.no_grad():
-        output = model(ids)
-    raw_logits = output.logits[0, -1].cpu().numpy()
-
-    # ----- dummy semantic vectors (demo only) -----
-    G = np.random.randn(256); G /= np.linalg.norm(G)
-    I = G + np.random.normal(scale=0.05, size=256)
-
-    # ----- run or skip WFGY -----
-    if enable_wfgy:
-        mod_logits = ENGINE.run(
-            input_vec=I,
-            ground_vec=G,
-            logits=raw_logits
+        return "", "", "<i>Please enter a prompt.</i>", None
+
+    try:
+        ids = tokenizer(prompt, return_tensors="pt").input_ids
+        raw_logits = model(ids).logits[0, -1].detach().numpy()
+
+        G = np.random.randn(256); G /= np.linalg.norm(G)
+        I = G + np.random.normal(scale=0.05, size=256)
+
+        mod_logits = (
+            ENGINE.run(input_vec=I, ground_vec=G, logits=raw_logits)
+            if enable_wfgy else raw_logits.copy()
         )
-    else:
-        mod_logits = raw_logits.copy()
-
-    # ----- decode 1-step continuation -----
-    raw_next = tokenizer.decode(int(raw_logits.argmax()))
-    mod_next = tokenizer.decode(int(mod_logits.argmax()))
-    raw_txt  = prompt + raw_next
-    mod_txt  = prompt + mod_next
-
-    # ----- metrics -----
-    m = compare_logits(raw_logits, mod_logits)
-    top1_flag = "✔ changed" if m["top1_shift"] else "✘ unchanged"
-    badge = (
-        f"<b>variance ▼ {(1-m['std_ratio'])*100:.0f}%</b> "
-        f"| <b>KL {m['kl_divergence']:.2f}</b> "
-        f"| top-1 {top1_flag}"
-    )
 
-    # ----- histogram -----
-    fig = plot_histogram(raw_logits, mod_logits, show=False)
-    buffer = io.BytesIO(); fig.savefig(buffer, format="png"); fig.clf()
-    hist_uri = "data:image/png;base64," + base64.b64encode(buffer.getvalue()).decode()
+        m = compare_logits(raw_logits, mod_logits)
+        top1 = "✔" if m["top1_shift"] else "✘"
+        metrics_html = (
+            f"<b>variance ▼ {(1-m['std_ratio'])*100:.0f}%</b> "
+            f"| <b>KL {m['kl_divergence']:.2f}</b> "
+            f"| top-1 {top1}"
+        )
 
-    return raw_txt, mod_txt, badge, hist_uri
+        fig = plot_histogram(raw_logits, mod_logits, show=False)
+        buf = io.BytesIO(); fig.savefig(buf, format="png"); fig.clf()
+        img_uri = "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode()
 
+        raw_next = tokenizer.decode(int(raw_logits.argmax()))
+        mod_next = tokenizer.decode(int(mod_logits.argmax()))
+        return prompt + raw_next, prompt + mod_next, metrics_html, img_uri
 
-# ------------------------------------------------------------
-# Gradio UI
-# ------------------------------------------------------------
-with gr.Blocks(title="WFGY — Self-Healing Variance Gate") as demo:
+    except Exception as e:
+        return "", "", f"<b style='color:red'>Error:</b> {str(e)}", None
 
+
+css = """
+#prompt-row {margin-bottom: 1.0rem}
+.gr-box {font-size: 0.85rem}
+"""
+
+with gr.Blocks(css=css, theme=gr.themes.Soft()) as demo:
     gr.Markdown(
         """
-### 🧠 **WFGY 1-click Variance Gate**
+### 🧠 WFGY 1-click Variance Gate  
 
-*Turn GPT-2 into a calmer thinker in seconds. Bigger LLMs show even stronger gains.*
+Turn GPT-2 into a calmer thinker in seconds.<br>
+**Bigger LLMs → even stronger gains.**
 
 | Metric | Meaning |
 |--------|---------|
-| **variance ▼** | logits become less noisy (focus ↑) |
-| **KL** | distribution actually reshaped |
-| **top-1** | most-likely token swapped ✔ or not ✘ |
+| **variance ▼** | logits become less noisy |
+| **KL** | distribution reshaped |
+| **top-1** | most-likely token swapped ✔/✘ |
 
-**Benchmarks (WFGY 1.0 vs base):**
+**Benchmarks (WFGY 1.0 vs base)**  
 
 | Task | Base % | WFGY % | Δ |
-|------|-------|-------|---|
+|------|-------|--------|---|
 | MMLU | 61.0 | **89.8** | +47 % |
 | TruthfulQA | 62.4 | **90.4** | +45 % |
 | GSM8K | 78.0 | **98.7** | +27 % |
-
-> 🔖 *PDF workflow*: clone repo → feed `docs/WFGY_1.0.pdf` to <em>any</em> chat-LLM, prepend your prompt with **“use WFGY”** and watch the difference — no-code, cross-model magic.  
-> ⭐ **10 000 GitHub Stars before 2025-08-01** unlocks **WFGY 2.0** (adaptive gamma, cross-modal). Miss it and v2 goes pay-walled & sealed.  
-> 📂 *I_am_not_lizardman/* holds <b>8 + 1 “Challenge-Einstein” papers</b> — tweet a screenshot if you find them!
         """
     )
 
-    with gr.Row():
-        prompt   = gr.Textbox(label="Prompt", placeholder="Ask anything…", lines=2)
-        enable   = gr.Checkbox(label="Enable WFGY", value=True)
-        run_btn  = gr.Button("Run")
+    with gr.Row(elem_id="prompt-row"):
+        prompt = gr.Textbox(label="Prompt", lines=2, placeholder="Ask anything…")
+        enable = gr.Checkbox(label="Enable WFGY", value=True)
+        run_btn = gr.Button("Run")
 
     with gr.Row():
-        raw_out  = gr.Textbox(label="• Raw GPT-2", lines=4)
-        mod_out  = gr.Textbox(label="• After WFGY", lines=4)
+        raw_box = gr.Textbox(label="Raw GPT-2")
+        mod_box = gr.Textbox(label="After WFGY")
 
     metrics = gr.HTML()
-    hist    = gr.Image(label="Logit distribution", width=440)
+    hist_img = gr.Image(label="Logit distribution", width=440)
 
-    run_btn.click(
-        fn=wfgy_demo,
-        inputs=[prompt, enable],
-        outputs=[raw_out, mod_out, metrics, hist]
-    )
+    run_btn.click(wfgy_pipeline, [prompt, enable],
+                  [raw_box, mod_box, metrics, hist_img])
 
     gr.Markdown(
         """
-<div align="center">
-⭐ Love the variance drop? <a href="https://github.com/onestardao/WFGY" target="_blank"><b>Star the repo</b></a> &nbsp;•&nbsp;
-<a href="https://doi.org/10.5281/zenodo.15630970" target="_blank">Read the paper</a>
-</div>
+**PDF mode** – feed <code>I_am_not_lizardman/WFGY_1.0.pdf</code> to any chat-LLM,  
+prepend <code>Use WFGY:</code> and watch replies get sharper. Prompt revolution!  
+
+⭐ **10 000 GitHub stars before 2025-08-01** unlocks **WFGY 2.0**  
+(secret adaptive-gamma, multimodal edition).  
+
+📂 Hidden folder <b>I_am_not_lizardman/</b> holds 8 + 1 “Challenge-Einstein” papers — tweet a screenshot if you find them!
         """,
         elem_id="footer"
     )