Create app.py

app.py

@@ -0,0 +1,90 @@
+import gradio as gr
+import torch
+import html as html_lib
+from transformers import AutoTokenizer, AutoModelForCausalLM
+
+tokenizer = AutoTokenizer.from_pretrained("BigSalmon/InformalToFormalLincoln123Paraphrase")
+model = AutoModelForCausalLM.from_pretrained("BigSalmon/InformalToFormalLincoln123Paraphrase")
+model.eval()
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+model.to(device)
+
+def get_color(p):
+    hue = min(p * 120, 120)
+    return f"hsl({hue},80%,35%)", f"hsla({hue},80%,50%,0.15)"
+
+def analyze_text(text, top_k):
+    top_k = max(1, int(top_k))
+    if not text.strip():
+        return "<p style='color:#999;text-align:center;padding:40px'>Paste some text and click Analyze.</p>"
+
+    tokens = tokenizer.encode(text)
+    if len(tokens) > 512:
+        tokens = tokens[:512]
+
+    with torch.no_grad():
+        input_ids = torch.tensor([tokens]).to(device)
+        all_logits = model(input_ids).logits[0].cpu()
+
+    css = """<style>
+.tc{display:flex;flex-wrap:wrap;gap:5px;padding:20px;line-height:2.4;font-family:'Segoe UI',sans-serif}
+.tw{position:relative;display:inline-block}
+.tk{padding:4px 7px;border-radius:6px;cursor:default;font-size:15px;transition:.2s;border:1px solid transparent}
+.tw:hover .tk{transform:translateY(-2px);box-shadow:0 4px 14px rgba(0,0,0,.18);border-color:#999}
+.tt{display:none;position:absolute;bottom:calc(100% + 8px);left:50%;transform:translateX(-50%);
+background:#1a1a2e;color:#eee;padding:14px;border-radius:12px;font-size:13px;z-index:9999;
+box-shadow:0 10px 30px rgba(0,0,0,.35);min-width:220px;max-height:350px;overflow-y:auto}
+.tw:hover .tt{display:block}
+.th{font-weight:700;font-size:14px;color:#7fdbca;border-bottom:1px solid #333;padding-bottom:6px;margin-bottom:6px}
+.tp{color:#ffd700;margin-bottom:8px}
+.at{color:#ff79c6;font-size:10px;text-transform:uppercase;letter-spacing:1px;margin-bottom:4px}
+.aw{display:flex;justify-content:space-between;padding:2px 0;font-size:12px}
+.aw .w{color:#c3cee3}.aw .p{color:#666;margin-left:14px}
+.hi{font-weight:700;color:#7fdbca!important}
+</style>"""
+
+    parts = [css, '<div class="tc">']
+    for i in range(len(tokens)):
+        tok = html_lib.escape(tokenizer.decode([tokens[i]]))
+        if i == 0:
+            parts.append(f'<div class="tw"><span class="tk" style="background:rgba(128,128,128,.1);color:#888">{tok}</span></div>')
+            continue
+
+        probs = torch.softmax(all_logits[i - 1], dim=-1)
+        actual_p = probs[tokens[i]].item()
+        top_p, top_idx = probs.topk(top_k)
+        color, bg = get_color(actual_p)
+
+        rank = None
+        alts = ""
+        for j in range(top_k):
+            a_text = html_lib.escape(tokenizer.decode([top_idx[j].item()]))
+            a_p = top_p[j].item()
+            hit = top_idx[j].item() == tokens[i]
+            if hit: rank = j + 1
+            cls = ' class="w hi"' if hit else ' class="w"'
+            pcls = ' class="p hi"' if hit else ' class="p"'
+            alts += f'<div class="aw"><span{cls}>{a_text}</span><span{pcls}>{a_p:.4f}</span></div>'
+
+        rank_s = f"rank #{rank}" if rank else f"rank &gt;{top_k}"
+        tooltip = f'''<div class="tt">
+<div class="th">&ldquo;{tok}&rdquo;</div>
+<div class="tp">P = {actual_p:.4f} &nbsp;({rank_s})</div>
+<div class="at">Top {top_k} alternatives</div>{alts}</div>'''
+
+        parts.append(f'<div class="tw"><span class="tk" style="background:{bg};color:{color}">{tok}</span>{tooltip}</div>')
+
+    parts.append('</div>')
+    return ''.join(parts)
+
+
+with gr.Blocks(theme=gr.themes.Soft(), css="footer{display:none!important}.main{max-width:960px;margin:auto}") as demo:
+    gr.Markdown("# 🔍 Token Probability Explorer\nPaste text, hover over each token to see its probability and the most likely alternatives.")
+    with gr.Row():
+        text_input = gr.Textbox(label="Input Text", placeholder="Paste your text here…", lines=5, scale=4)
+        top_k_input = gr.Number(label="# Alternatives", value=10, minimum=1, maximum=200, step=1, scale=1)
+    btn = gr.Button("Analyze", variant="primary")
+    output = gr.HTML()
+    btn.click(fn=analyze_text, inputs=[text_input, top_k_input], outputs=output)
+
+demo.launch()