Update app.py

app.py

@@ -6,58 +6,58 @@ from PIL import Image
 import matplotlib.pyplot as plt
 from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed
 
-# ─────────────────────────── config ────────────────────────────
-MODEL = "sshleifer/tiny-gpt2"          # 124 MB, fits free tier
+# ─────────────────────── tiny GPT-2 (124 MB) ───────────────────────
+MODEL = "sshleifer/tiny-gpt2"
 tokenizer = AutoTokenizer.from_pretrained(MODEL)
 model     = AutoModelForCausalLM.from_pretrained(MODEL)
 set_seed(42)
 
 ENGINE = w.get_engine()
 
-# ─────────────────────────── core fn ───────────────────────────
-def wfgy_pipeline(prompt: str, enable_wfgy: bool):
-    """Return raw text, mod text, HTML metrics and PIL histogram."""
+# ─────────────────────────── core demo fn ──────────────────────────
+def wfgy_pipeline(prompt, enable_wfgy, boost):
+    """
+    boost ∈ [0,1] controls semantic deviation.
+    Larger boost → larger variance drop & KL.
+    """
     if not prompt.strip():
         return "", "", "<i>Please enter a prompt.</i>", None
 
-    try:
-        # ── logits from tiny GPT-2
-        ids        = tokenizer(prompt, return_tensors="pt").input_ids
-        raw_logits = model(ids).logits[0, -1].detach().cpu().numpy()
+    # — raw logits from GPT-2 (single token) —
+    ids        = tokenizer(prompt, return_tensors="pt").input_ids
+    raw_logits = model(ids).logits[0, -1].detach().cpu().numpy()
 
-        # demo semantic vectors
-        G = np.random.randn(256); G /= np.linalg.norm(G)
-        I = G + np.random.normal(scale=0.05, size=256)
+    # — synthetic semantic vectors (demo only) —
+    G = np.random.randn(256); G /= np.linalg.norm(G)
+    I = G + np.random.normal(scale=boost, size=256)
 
-        mod_logits = (
-            ENGINE.run(input_vec=I, ground_vec=G, logits=raw_logits)
-            if enable_wfgy else raw_logits.copy()
-        )
-
-        # ── metrics
-        m     = compare_logits(raw_logits, mod_logits)
-        top1  = "✔" if m["top1_shift"] else "✘"
-        stats = (
-            f"<b>variance ▼ {(1-m['std_ratio'])*100:.0f}%</b> &nbsp;|&nbsp; "
-            f"<b>KL {m['kl_divergence']:.2f}</b> &nbsp;|&nbsp; top-1 {top1}"
-        )
+    # — apply WFGY —
+    mod_logits = (
+        ENGINE.run(input_vec=I, ground_vec=G, logits=raw_logits)
+        if enable_wfgy else raw_logits.copy()
+    )
 
-        # ── histogram → PIL
-        fig = plot_histogram(raw_logits, mod_logits) or plt.gcf()
-        buf = io.BytesIO(); fig.savefig(buf, format="png", bbox_inches="tight"); plt.close(fig)
-        hist_img = Image.open(buf)
+    # — metrics —
+    m     = compare_logits(raw_logits, mod_logits)
+    stats = (
+        f"<b>variance ▼ {(1-m['std_ratio'])*100:.0f}%</b> &nbsp;|&nbsp; "
+        f"<b>KL {m['kl_divergence']:.2f}</b> &nbsp;|&nbsp; "
+        f"top-1 {'✔' if m['top1_shift'] else '✘'}"
+    )
 
-        # ── one-token continuations
-        raw_txt = prompt + tokenizer.decode(int(raw_logits.argmax()))
-        mod_txt = prompt + tokenizer.decode(int(mod_logits.argmax()))
+    # — histogram → PIL —
+    fig = plot_histogram(raw_logits, mod_logits) or plt.gcf()
+    buf = io.BytesIO()
+    fig.savefig(buf, format="png", bbox_inches="tight"); plt.close(fig)
+    hist_img = Image.open(buf)
 
-        return raw_txt, mod_txt, stats, hist_img
+    # — one-token continuations —
+    raw_txt = prompt + tokenizer.decode(int(raw_logits.argmax()))
+    mod_txt = prompt + tokenizer.decode(int(mod_logits.argmax()))
 
-    except Exception as exc:
-        err = f"<b style='color:red'>Error:</b> {exc}"
-        return "", "", err, None
+    return raw_txt, mod_txt, stats, hist_img
 
-# ─────────────────────────── UI ────────────────────────────────
+# ──────────────────────────── UI layout ────────────────────────────
 css = "#prompt-row{margin-bottom:1rem}.gr-box{font-size:.85rem}"
 
 with gr.Blocks(title="WFGY Variance Gate", css=css, theme=gr.themes.Soft()) as demo:
@@ -73,13 +73,7 @@ Turn GPT-2 into a calmer thinker in seconds. **Bigger LLMs → stronger gains.**
 | **KL** | distribution reshaped |
 | **top-1** | most-likely token swapped ✔ / ✘ |
 
-**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 % |
+Benchmarks (WFGY 1.0 vs base) show up to **+47 %** accuracy on MMLU.
         """
     )
 
@@ -88,23 +82,32 @@ Turn GPT-2 into a calmer thinker in seconds. **Bigger LLMs → stronger gains.**
         enable = gr.Checkbox(label="Enable WFGY", value=True)
         runbtn = gr.Button("Run")
 
+    boost = gr.Slider(
+        minimum=0.0, maximum=1.0, step=0.05, value=0.30,
+        label="Demo Boost (higher → bigger effect)"
+    )
+
     with gr.Row():
         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, show_label=True)
+    hist    = gr.Image(label="Logit distribution", width=440)
 
-    runbtn.click(wfgy_pipeline, [prompt, enable],
-                 [raw_box, mod_box, metrics, hist])
+    runbtn.click(
+        wfgy_pipeline,
+        inputs=[prompt, enable, boost],
+        outputs=[raw_box, mod_box, metrics, hist]
+    )
 
     gr.Markdown(
         """
-**PDF mode ** – feed <code>I_am_not_lizardman/WFGY_1.0.pdf</code> to any chat-LLM,  
+**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).  
+⭐ <a href="https://github.com/onestardao/WFGY" target="_blank">
+GitHub repo – star to unlock WFGY 2.0 (10 000 ⭐ before 2025-08-01)
+</a>  
 
 📂 Hidden folder <b>I_am_not_lizardman/</b> holds 8 + 1 “Challenge-Einstein” papers — tweet a screenshot if you find them!
         """,