| import spaces |
| import gradio as gr |
| import torch |
| from transformers import AutoTokenizer, AutoModelForCausalLM |
| from peft import PeftModel |
|
|
| base_model_id = "unsloth/llama-3-8b-Instruct-bnb-4bit" |
| adapter_model_id = "mutij8/academic-humanizer-llama3" |
|
|
| tokenizer = None |
| model = None |
|
|
| def load_model_once(): |
| global tokenizer, model |
| if model is not None: |
| return tokenizer, model |
|
|
| print("⏳ ZeroGPU allocated! Loading base model and fine-tuned adapter...") |
| tokenizer = AutoTokenizer.from_pretrained(adapter_model_id) |
| |
| base_model = AutoModelForCausalLM.from_pretrained( |
| base_model_id, |
| torch_dtype=torch.float16, |
| device_map="cuda" |
| ) |
| model = PeftModel.from_pretrained(base_model, adapter_model_id) |
| print("✅ Model loaded successfully on GPU!") |
| return tokenizer, model |
|
|
| @spaces.GPU(duration=120) |
| def humanize_and_correct(paragraph): |
| paragraph = paragraph.strip() |
| if not paragraph: |
| return "Please enter a valid academic paragraph." |
|
|
| tok, mod = load_model_once() |
| word_count = len(paragraph.split()) |
|
|
| dual_instruction = f"""You are an expert academic editor and researcher. Your task is to simultaneously fix all grammatical errors, spelling mistakes, and awkward phrasing while humanizing the text. |
| - Tone: Write with natural human imperfection, varied sentence lengths, and stylistic asymmetry. |
| - Rule 1: Output MUST be between {int(word_count*0.85)} and {int(word_count*1.15)} words. |
| - Rule 2: Fix any grammar, punctuation, tense inconsistencies, or syntax errors present in the input text. |
| - Rule 3: Force clause inversion and structural variety to eliminate AI detection signatures and ensure low plagiarism. |
| - Rule 4: Keep all inline citations (e.g. [1] or (Smith, 2023)) 100% intact and correctly placed. |
| - Rule 5: Output ONLY the fully corrected and humanized text.""" |
|
|
| messages = [ |
| {"role": "system", "content": dual_instruction}, |
| {"role": "user", "content": paragraph} |
| ] |
|
|
| prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) |
| inputs = tok([prompt], return_tensors="pt").to("cuda") |
|
|
| outputs = mod.generate( |
| **inputs, |
| max_new_tokens=int(word_count * 1.5) + 30, |
| do_sample=True, |
| temperature=0.85, |
| top_p=0.95, |
| top_k=50 |
| ) |
|
|
| input_len = inputs["input_ids"].shape[1] |
| generated_text = tok.decode(outputs[0][input_len:], skip_special_tokens=True).strip() |
| return generated_text |
|
|
| interface = gr.Interface( |
| fn=humanize_and_correct, |
| inputs=gr.Textbox(lines=8, placeholder="Paste your text with grammar errors or AI style here..."), |
| outputs=gr.Textbox(lines=8, label="Grammar-Corrected & Humanized Output"), |
| title="Turnitin-Calibrated Dual-Action Academic Humanizer", |
| description="Simultaneously fixes grammatical errors and transforms AI or flawed text into Turnitin-verified human prose." |
| ) |
|
|
| if __name__ == "__main__": |
| interface.launch() |
