ToxicGPT

About ToxicGPT

ToxicGPT began as an exploration into bias and toxicity in language models, and evolved into a hands-on deep dive into alignment, lightweight fine-tuning, and ethical deployment. We fine-tuned GPT-2 on the Jigsaw Toxic Comment dataset using QLoRA and PEFT (Parameter-Efficient Fine-Tuning), achieving faster training with minimal compute.

Once trained, the model was deployed via a Flask backend with session tracking and exposed to a React.js frontend that allowed users to experiment with toxicity prompts. To push the boundaries of the project further, we integrated Stable Diffusion to generate multimodal outputs (image generation based on text prompts), adding a fascinating visual layer to toxic or aligned text.

• Efficient Fine-Tuning with QLoRA
  Applied 4-bit quantization and LoRA adapters to fine-tune GPT-2 on toxic text using minimal resources.

• Flask API & Session Logging
  Built a fast, lightweight REST API with per-session tracking for input prompts, outputs, and feedback.

• Interactive React Frontend
  Developed an intuitive interface with prompt input, generation preview.

• MLflow for Experiment Tracking
  All hyperparameters, model versions, and metrics logged using MLflow for reproducibility and comparison.

• DVC for Dataset & Model Versioning
  Managed large model checkpoints and the Jigsaw dataset efficiently with DVC for clean collaboration and history.

Tech Stack

• Model & Training: Transformers, PEFT, QLoRA
• Experiment Tracking: MLflow, DVC
• Deployment: Flask
• Frontend: React.js, Tailwind CSS, shadcn/ui
• Multimodal: Stable Diffusion (via diffusers), Hugging Face pipelines
• Infrastructure: Google Colab + GPU, GitHub Actions for CI/CD

Credits

• Hugging Face Transformers & Datasets: For fast integration of pretrained models and data.
• MLflow: Experiment tracking with ease.
• DVC: Seamless version control for datasets and model weights.
• Stable Diffusion: Generative backbone for visualizing text in image form.

Author

Built by Rohit Kshirsagar & Rishabh Kothari, founding engineer’s at ApexAI and AI systems enthusiast.
Check out more projects on GitHub or reach out via LinkedIn.