Update app.py

app.py

@@ -1,146 +1,142 @@
-import gradio as gr
-import torch
-from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
-from transformers.pipelines import pipeline
-from sentence_transformers import SentenceTransformer, util
-import numpy as np
-import gradio.themes as grthemes
-
-# Paraphrasing model: humarin/chatgpt_paraphraser_on_T5_base
-PARAPHRASE_MODEL_NAME = "humarin/chatgpt_paraphraser_on_T5_base"
-paraphrase_tokenizer = AutoTokenizer.from_pretrained(PARAPHRASE_MODEL_NAME)
-paraphrase_model = AutoModelForSeq2SeqLM.from_pretrained(PARAPHRASE_MODEL_NAME)
-device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
-paraphrase_model = paraphrase_model.to(device)
-
-# AI Detector: roberta-base-openai-detector
-ai_detector = pipeline("text-classification", model="roberta-base-openai-detector", device=0 if torch.cuda.is_available() else -1)
-
-# Semantic similarity model
-similarity_model = SentenceTransformer('all-MiniLM-L6-v2')
-
-tone_templates = {
-    "Academic": "Paraphrase the following text in a formal, academic tone:",
-    "Casual": "Paraphrase the following text in a casual, conversational tone:",
-    "Friendly": "Paraphrase the following text in a friendly, approachable tone:",
-    "Stealth": "Paraphrase the following text to bypass AI detectors and sound as human as possible:",
-}
-
-# Paraphrasing function
-def paraphrase(text, tone):
-    prompt = tone_templates[tone] + " " + text
-    input_ids = paraphrase_tokenizer(
-        f'paraphrase: {prompt}',
-        return_tensors="pt", padding="longest",
-        max_length=256, truncation=True
-    ).input_ids.to(device)
-    outputs = paraphrase_model.generate(
-        input_ids,
-        temperature=0.7,
-        repetition_penalty=1.2,
-        num_return_sequences=1,
-        no_repeat_ngram_size=2,
-        max_length=256,
-        diversity_penalty=3.0,
-        num_beams=5,
-        num_beam_groups=5
-    )
-    res = paraphrase_tokenizer.batch_decode(outputs, skip_special_tokens=True)
-    return res[0] if res else ""
-
-def semantic_similarity(text1, text2):
-    emb1 = similarity_model.encode(text1, convert_to_tensor=True)
-    emb2 = similarity_model.encode(text2, convert_to_tensor=True)
-    sim = util.pytorch_cos_sim(emb1, emb2).item()
-    return sim
-
-def ai_detect(text):
-    # Returns probability of being AI-generated (label 'Fake')
-    result = ai_detector(text)
-    for r in result:
-        if r['label'] == 'Fake':
-            return r['score']
-        elif r['label'] == 'Real':
-            return 1.0 - r['score']
-    return 0.5  # fallback
-
-def humanization_score(sim, ai_prob):
-    # Lower similarity and lower AI probability = more human
-    score = (1.0 - sim) * 0.5 + (1.0 - ai_prob) * 0.5
-    return score
-
-def humanization_rating(score):
-    if score < 0.7:
-        return f"⚠️ Still robotic ({score:.2f})"
-    elif score < 0.85:
-        return f"👍 Acceptable ({score:.2f})"
-    else:
-        return f"✅ Highly Human ({score:.2f})"
-
-def process(text, tone):
-    if not text.strip():
-        return "", 0.0, 0.0, 0.0, "", 0.0, ""
-    # Pre-humanization AI detection
-    pre_ai_prob = ai_detect(text)
-    # Paraphrase
-    paraphrased = paraphrase(text, tone)
-    # Post-humanization AI detection
-    post_ai_prob = ai_detect(paraphrased)
-    # Semantic similarity
-    sim = semantic_similarity(text, paraphrased)
-    # Humanization score
-    score = humanization_score(sim, post_ai_prob)
-    rating = humanization_rating(score)
-    return (
-        paraphrased,
-        pre_ai_prob * 100,
-        post_ai_prob * 100,
-        sim,
-        rating,
-        score * 100,
-        f"Pre: {pre_ai_prob*100:.1f}% | Post: {post_ai_prob*100:.1f}%"
-    )
-
-# Custom dark theme using gradio.themes.Base
-custom_theme = grthemes.Base(
-    primary_hue="blue",
-    secondary_hue="blue",
-    neutral_hue="slate"
-)
-
-with gr.Blocks(theme=custom_theme) as demo:
-    gr.Markdown("""
-    # 🧠 AI Humanizer
-    <div style='display:flex;justify-content:space-between;align-items:center;'>
-        <span style='font-size:1.2em;color:#7bb1ff;'>Rewrite AI text to sound 100% human</span>
-        <span style='font-weight:bold;color:#7bb1ff;'>Made by Taha</span>
-    </div>
-    """, elem_id="header")
-    with gr.Row():
-        with gr.Column():
-            text_in = gr.Textbox(label="Paste AI-generated text here", lines=8, placeholder="Paste your text...")
-            tone = gr.Radio(["Academic", "Casual", "Friendly", "Stealth"], value="Stealth", label="Tone Selector")
-            btn = gr.Button("Humanize", elem_id="humanize-btn")
-        with gr.Column():
-            text_out = gr.Textbox(label="Humanized Output", lines=8, interactive=False)
-            rating = gr.Markdown("", elem_id="rating")
-            gr.Markdown("<hr/>")
-            gr.Markdown("<b>AI Probability Scores</b>")
-            ai_scores = gr.Markdown("", elem_id="ai-scores")
-            gr.Markdown("<b>Semantic Similarity</b>")
-            sim_score = gr.Number(label="Similarity (0=very different, 1=very similar)", interactive=False)
-            gr.Markdown("<b>Humanization %</b>")
-            human_score = gr.Number(label="Humanization Score (%)", interactive=False)
-    btn.click(
-        process,
-        inputs=[text_in, tone],
-        outputs=[text_out, ai_scores, ai_scores, sim_score, rating, human_score, ai_scores],
-        api_name="humanize"
-    )
-    gr.Markdown("""
-    <div style='text-align:center;color:#7bb1ff;margin-top:2em;'>
-        <b>Made by Taha</b> | Free for unlimited use | Optimized for students
-    </div>
-    """, elem_id="footer")
-
+import gradio as gr
+import torch
+from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+from transformers.pipelines import pipeline
+from sentence_transformers import SentenceTransformer, util
+import numpy as np
+import gradio.themes as grthemes
+
+# Paraphrasing model: humarin/chatgpt_paraphraser_on_T5_base
+PARAPHRASE_MODEL_NAME = "humarin/chatgpt_paraphraser_on_T5_base"
+paraphrase_tokenizer = AutoTokenizer.from_pretrained(PARAPHRASE_MODEL_NAME)
+paraphrase_model = AutoModelForSeq2SeqLM.from_pretrained(PARAPHRASE_MODEL_NAME)
+device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+paraphrase_model = paraphrase_model.to(device)
+
+# AI Detector: roberta-base-openai-detector
+ai_detector = pipeline("text-classification", model="roberta-base-openai-detector", device=0 if torch.cuda.is_available() else -1)
+
+# Semantic similarity model
+similarity_model = SentenceTransformer('all-MiniLM-L6-v2')
+
+tone_templates = {
+    "Academic": "Paraphrase the following text in a formal, academic tone:",
+    "Casual": "Paraphrase the following text in a casual, conversational tone:",
+    "Friendly": "Paraphrase the following text in a friendly, approachable tone:",
+    "Stealth": "Paraphrase the following text to bypass AI detectors and sound as human as possible:",
+}
+
+# Paraphrasing function
+def paraphrase(text, tone):
+    prompt = tone_templates[tone] + " " + text
+    input_ids = paraphrase_tokenizer(
+        f'paraphrase: {prompt}',
+        return_tensors="pt", padding="longest",
+        max_length=256, truncation=True
+    ).input_ids.to(device)
+    outputs = paraphrase_model.generate(
+        input_ids,
+        temperature=0.7,
+        repetition_penalty=1.2,
+        num_return_sequences=1,
+        no_repeat_ngram_size=2,
+        max_length=256,
+        diversity_penalty=3.0,
+        num_beams=5,
+        num_beam_groups=5,
+        trust_remote_code=True
+    )
+    res = paraphrase_tokenizer.batch_decode(outputs, skip_special_tokens=True)
+    return res[0] if res else ""
+
+def semantic_similarity(text1, text2):
+    emb1 = similarity_model.encode(text1, convert_to_tensor=True)
+    emb2 = similarity_model.encode(text2, convert_to_tensor=True)
+    sim = util.pytorch_cos_sim(emb1, emb2).item()
+    return sim
+
+def ai_detect(text):
+    # Returns probability of being AI-generated (label 'Fake')
+    result = ai_detector(text)
+    for r in result:
+        if r['label'] == 'Fake':
+            return r['score']
+        elif r['label'] == 'Real':
+            return 1.0 - r['score']
+    return 0.5  # fallback
+
+def humanization_score(sim, ai_prob):
+    # Lower similarity and lower AI probability = more human
+    score = (1.0 - sim) * 0.5 + (1.0 - ai_prob) * 0.5
+    return score
+
+def humanization_rating(score):
+    if score < 0.7:
+        return f"⚠️ Still robotic ({score:.2f})"
+    elif score < 0.85:
+        return f"👍 Acceptable ({score:.2f})"
+    else:
+        return f"✅ Highly Human ({score:.2f})"
+
+def process(text, tone):
+    if not text.strip():
+        return "", "", 0.0, "", 0.0
+    # Pre-humanization AI detection
+    pre_ai_prob = ai_detect(text)
+    # Paraphrase
+    paraphrased = paraphrase(text, tone)
+    # Post-humanization AI detection
+    post_ai_prob = ai_detect(paraphrased)
+    # Semantic similarity
+    sim = semantic_similarity(text, paraphrased)
+    # Humanization score
+    score = humanization_score(sim, post_ai_prob)
+    rating = humanization_rating(score)
+    ai_score_str = f"Pre: {pre_ai_prob*100:.1f}% | Post: {post_ai_prob*100:.1f}%"
+    return (
+        paraphrased,         # gr.Textbox (string)
+        ai_score_str,        # gr.Markdown (string)
+        sim,                # gr.Number (float)
+        rating,             # gr.Markdown (string)
+        score * 100         # gr.Number (float)
+    )
+
+# Custom dark theme using gradio.themes.Base
+custom_theme = grthemes.Base(
+    primary_hue="blue",
+    secondary_hue="blue",
+    neutral_hue="slate"
+)
+
+with gr.Blocks(theme=custom_theme) as demo:
+    gr.Markdown("""
+    # 🧠 AI Humanizer
+    <div style='display:flex;justify-content:space-between;align-items:center;'>
+        <span style='font-size:1.2em;color:#7bb1ff;'>Rewrite AI text to sound 100% human</span>
+        <span style='font-weight:bold;color:#7bb1ff;'>Made by Taha</span>
+    </div>
+    """, elem_id="header")
+    with gr.Row():
+        with gr.Column():
+            text_in = gr.Textbox(label="Paste AI-generated text here", lines=8, placeholder="Paste your text...")
+            tone = gr.Radio(["Academic", "Casual", "Friendly", "Stealth"], value="Stealth", label="Tone Selector")
+            btn = gr.Button("Humanize", elem_id="humanize-btn")
+        with gr.Column():
+            text_out = gr.Textbox(label="Humanized Output", lines=8, interactive=False)
+            ai_scores = gr.Markdown("", elem_id="ai-scores")
+            sim_score = gr.Number(label="Similarity (0=very different, 1=very similar)", interactive=False)
+            rating = gr.Markdown("", elem_id="rating")
+            human_score = gr.Number(label="Humanization Score (%)", interactive=False)
+    btn.click(
+        process,
+        inputs=[text_in, tone],
+        outputs=[text_out, ai_scores, sim_score, rating, human_score],
+        api_name="humanize"
+    )
+    gr.Markdown("""
+    <div style='text-align:center;color:#7bb1ff;margin-top:2em;'>
+        <b>Made by Taha</b> | Free for unlimited use | Optimized for students
+    </div>
+    """, elem_id="footer")
+
 demo.launch()