import gradio as gr from transformers import AutoTokenizer, AutoModelForSequenceClassification import torch import numpy as np import matplotlib.pyplot as plt import io import pandas as pd from PIL import Image import gc import shap # 1. MODEL CONFIGURATION MODELS = { "MuRIL": "SpamX/MuRIL_V2", "XLM-RoBERTa": "SpamX/XLM_RoBERTa_V2" } cache = {"model": None, "tokenizer": None, "current_id": None, "last_winner": "MuRIL"} def purge_memory(): global cache cache.update({"model": None, "tokenizer": None, "current_id": None}) gc.collect() if torch.cuda.is_available(): torch.cuda.empty_cache() def get_assets(model_name): global cache model_id = MODELS[model_name] if cache["current_id"] != model_id: purge_memory() cache["tokenizer"] = AutoTokenizer.from_pretrained(model_id) cache["model"] = AutoModelForSequenceClassification.from_pretrained(model_id, low_cpu_mem_usage=True) cache["current_id"] = model_id return cache["tokenizer"], cache["model"] def get_gating_prediction(text): tk_m, md_m = get_assets("MuRIL") with torch.no_grad(): out_m = torch.softmax(md_m(**tk_m(text, return_tensors="pt")).logits, dim=1)[0] prob_m_spam = out_m[1].item() conf_m = abs(prob_m_spam - 0.5) * 2 tk_x, md_x = get_assets("XLM-RoBERTa") with torch.no_grad(): out_x = torch.softmax(md_x(**tk_x(text, return_tensors="pt")).logits, dim=1)[0] prob_x_spam = out_x[1].item() conf_x = abs(prob_x_spam - 0.5) * 2 if conf_m >= conf_x: final_prob_spam, winner = prob_m_spam, "MuRIL" else: final_prob_spam, winner = prob_x_spam, "XLM-RoBERTa" cache["last_winner"] = winner label = "SPAM" if final_prob_spam > 0.5 else "HAM" score = final_prob_spam if label == "SPAM" else 1 - final_prob_spam return label, score, winner def generate_shap_plot(tokens, shap_values, model_name): data = [] for t, v in zip(tokens, shap_values): clean_t = t.replace(' ', '').replace('▁', '') if clean_t not in ['[CLS]', '[SEP]', '[PAD]', '~~', '~~', '']: data.append((clean_t, v)) if not data: return None labels, values = zip(*data) colors = ['#ff4444' if v > 0 else '#00ff88' for v in values] plt_height = max(5, len(labels) * 0.4) plt.figure(figsize=(12, plt_height), facecolor='#050505') ax = plt.axes() ax.set_facecolor('#050505') ax.barh(labels, values, color=colors) ax.tick_params(axis='both', colors='white', labelsize=10) for s in ax.spines.values(): s.set_color('#333') plt.title(f"SHAP ATTRIBUTION: {model_name}", color='white', pad=20, fontsize=14) plt.xlabel("Shapley Value (Marginal Contribution)", color='white') plt.gca().invert_yaxis() buf = io.BytesIO() plt.savefig(buf, format='png', bbox_inches='tight', facecolor='#050505') buf.seek(0); plt.close() return Image.open(buf) def predict(text, model_choice): if not text.strip(): return {"NULL": 0}, "STATUS: INPUT_REQUIRED" try: if model_choice == "Confidence-based Gating": label, conf, winner = get_gating_prediction(text) info = f"Decision Authority: {winner}" else: tk, md = get_assets(model_choice) cache["last_winner"] = model_choice with torch.no_grad(): out = torch.softmax(md(**tk(text, return_tensors="pt")).logits, dim=1)[0] prob_spam = out[1].item() label = "SPAM" if prob_spam > 0.5 else "HAM" conf = prob_spam if label == "SPAM" else 1 - prob_spam info = f"Model: {model_choice}" return {label: conf, ("HAM" if label == "SPAM" else "SPAM"): 1-conf}, f"RESULT: {label} | {info}" except Exception as e: purge_memory() return {"ERROR": 1}, f"Error: {str(e)}" def explain(text, model_choice): """Formal SHAP Audit: Synchronized with Gating Result.""" if not text.strip(): return None, "### INPUT REQUIRED" # Force synchronization with the winner of the scan explainer_id = cache.get("last_winner", "MuRIL") if model_choice == "Confidence-based Gating" else model_choice tk, md = get_assets(explainer_id) def f(x): tv = tk(x.tolist(), padding=True, truncation=True, return_tensors="pt") with torch.no_grad(): outputs = md(**tv).logits return torch.softmax(outputs, dim=1).numpy() # Use a smaller perturbation sample for very short text like "Hello" explainer = shap.Explainer(f, tk) shap_results = explainer([text]) prediction_scores = f(np.array([text]))[0] target_idx = np.argmax(prediction_scores) verdict = "SPAM" if target_idx == 1 else "HAM" v_color = "#ff4444" if verdict == "SPAM" else "#00ff88" # Get values and tokens current_values = shap_results.values[0][:, target_idx] raw_tokens = tk.convert_ids_to_tokens(tk(text)["input_ids"]) expected_value = shap_results.base_values[0][target_idx] pos_impact, neg_impact, rows = 0, 0, [] # Fixed Loop: Use zip to ensure we don't go out of bounds for t, v in zip(raw_tokens, current_values): # Clean special characters based on the specific model's vocabulary clean_t = t.replace(' ', '').replace('▁', '').replace('Ġ', '') # Skip padding and special markers if clean_t in ['[CLS]', '[SEP]', '~~', '~~', '', ''] or not clean_t.strip(): continue if v > 0: pos_impact += v c = v_color nature = verdict else: neg_impact += v c = "#ff4444" if verdict == "HAM" else "#00ff88" nature = "SPAM" if verdict == "HAM" else "HAM" rows.append(f"{clean_t}{nature}{v:+.4f}") # Final Convergence Calculation final_val = expected_value + pos_impact + neg_impact net_shift = pos_impact + neg_impact # If no tokens were meaningful (very common for "Hello"), provide a fallback if not rows: rows.append(f"Neutral content: No significant feature attribution found.") report_html = f"""

SPAMX EXPLAINABILITY REPORT: {verdict}

MATHEMATICAL AUDIT ({explainer_id})

Audit initialized at a BASELINE of {expected_value:.4f}.

CALCULATION LOG:
  Baseline: {expected_value:.4f}
+ HAM IMPACT: { (pos_impact if verdict == "HAM" else neg_impact):+.4f}
+ SPAM IMPACT: { (neg_impact if verdict == "HAM" else pos_impact):+.4f}

                    = FINAL CONVERGENCE: {final_val:.4f} ({final_val:.2%})
                

Conclusion: A net shift of {net_shift:+.4f} confirmed the {verdict} verdict.

TOKEN-LEVEL SHAPLEY VALUES

{"".join(rows)}

FEATURE (TOKEN)	FEATURE INFLUENCE	SHAPLEY VALUE

""" # Generate Plot plot_val = current_values if target_idx == 1 else -current_values plot_img = generate_shap_plot(raw_tokens, plot_val, explainer_id) return plot_img, report_html # UI CONFIGURATION css = """ body, .gradio-container { background-color: #050505 !important; color: white !important; font-family: 'Courier New', monospace; } #title { text-align: center; color: #ff0000; letter-spacing: 5px; font-weight: 900; } .gr-button-primary { background: #aa0000 !important; border: none !important; } """ with gr.Blocks(css=css) as demo: gr.Markdown("# SPAMX", elem_id="title") with gr.Row(): with gr.Column(scale=2): input_text = gr.Textbox(label="INPUT", placeholder="Enter comment...", lines=4) model_select = gr.Dropdown(["MuRIL", "XLM-RoBERTa", "Confidence-based Gating"], value="Confidence-based Gating", label="CORE SELECTION") with gr.Row(): btn_run = gr.Button("SCAN", variant="primary") btn_exp = gr.Button("EXPLAIN", variant="secondary") with gr.Column(scale=1): output_metric = gr.Label(label="PROBABILITY") output_verdict = gr.Markdown("STATUS: READY", elem_id="output-box") with gr.Tabs(): with gr.TabItem("1. NARRATIVE"): explain_html = gr.HTML("

Analysis required.

") with gr.TabItem("2. VISUAL"): explain_plot = gr.Image(label="SHAP Attribution Map") gr.Markdown("---") gr.Examples( examples=[ ["BROOO!! SUBSCRIBE CHEYYATHAVAR SUBSCRIBE CHEYYU!!"], ["Sathyam parayalo video super! Pakshe check the link in my profile for cash."], ["Amazing production! Does anyone follow @Cr_ypto_Ma_ster strategy?🚀"], ["Don't click that link guys, it's a scam! Stay safe!"], ["Thakarppan video aliya! Waiting for next part!"] ], inputs=input_text, label="EXAMPLES" ) btn_run.click(predict, [input_text, model_select], [output_metric, output_verdict]) btn_exp.click(explain, [input_text, model_select], [explain_plot, explain_html]) if __name__ == "__main__": demo.launch()