Subhan5622/Humanize-Ai

Humanize AI by subhan.tech

Developed by: Subhan Haider
Website: subhan.tech
Model Type: Causal Language Model (LoRA Fine-tuned)
Base Model: facebook/opt-1.3b
Language: English
License: MIT

Model Description

The Humanize AI is an advanced language model fine-tuned specifically to rewrite and restructure AI-generated text so that it reads naturally and bypasses AI detection systems.

Developed by Subhan Haider, this model was trained locally on an RTX 4060 using a massive dataset of over 141,000 paired text samples spanning multiple domains (Reddit, Wikipedia, ELI5).

Training Architecture

The training pipeline uses a two-stage approach to align the model with human writing patterns:

graph TD
    A[Raw Datasets<br/>Reddit, ELI5, Wikipedia] -->|Extract Human Text| B(Dataset Builder)
    B --> C[136k Human Text Samples]
    C -->|Feed to OPT-125m| D(Synthetic AI Generator)
    D --> E[141k AI vs Human Text Pairs]
    
    E --> F{Stage 1: Supervised Fine-Tuning}
    F -->|Learn Rewriting Patterns| G[SFT Model Checkpoint]
    
    G --> H{Stage 2: Direct Preference Optimization}
    H -->|RoBERTa AI Detector Scoring| I[Chosen vs Rejected Pairs]
    I --> J[Final AI Humanizer Model]
    
    J --> K[FastAPI Server & Web UI]

Stage 1: Supervised Fine-Tuning (SFT)

The model was first trained to convert stiff, robotic AI outputs into fluid, human-like responses. It learned vocabulary variance, sentence length pacing, and organic formatting.

Stage 2: Direct Preference Optimization (DPO)

The model was aligned against a roberta-base-openai-detector. The detector scored generated texts, separating them into "Chosen" (bypassed detection) and "Rejected" (flagged as AI). DPO actively penalized the "AI-sounding" patterns and rewarded organic human quirks.

Training Details

• Hardware: NVIDIA RTX 4060 (8GB VRAM)
• Quantization: 8-bit (bitsandbytes)
• Adapter: LoRA (Low-Rank Adaptation)
• Effective Batch Size: 32
• Learning Rate: 2e-4 (SFT), 5e-5 (DPO)

Intended Use

This model is designed to assist writers, marketers, and students in naturalizing machine-generated content. You can run the model locally using the built-in FastAPI server and Web UI.

Quick Start

To run the humanizer dashboard locally, execute the master menu and select the UI option:

run.bat

Then navigate to http://localhost:8000 in your browser.

Dataset Sources

The human-side data was synthetically paired against OPT-125m outputs to create a robust contrast. Human samples were sourced from:

• eli5_category
• reddit_tifu
• writing_prompts
• wikipedia

---

License

MIT License

Copyright (c) 2026 Subhan Haider

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.