AI that generates Amazon-style product descriptions using a fine-tuned GPT-Neo model. Built as a proof of concept for Texta.ai.
The founder of Texta.ai reached out — they were using GPT-3 to generate product descriptions but it was too expensive. I built a cheaper alternative by fine-tuning GPT-Neo on scraped Amazon product data.
Started with Puppeteer scraping from Amazon directly, but it was painfully slow. Switched to a pre-scraped dataset from UCSD. Cleaned the data, shaped it into prompt templates, and fine-tuned multiple models using HappyTransformer on Google Colab.
There are 4 model variants (0–3), each trained progressively longer. Model 3 gives the best results.
pip install happytransformerThen open the notebook:
jupyter notebook Texta_ai_Amazon_gpt_neo.ipynbDownload models from Google Drive or use the ones in this repo. Point load_path to your chosen model directory.
Video tutorial (in Russian)