Girish Kurup
10 min readMay 26, 2023

--

How was GPT 3 trained by OpenAI? How to train GPT 3 on Data, content specific to my company?

Photo by Andrew Neel on Unsplash

Open AI trained GPT 3 .5 model ( pre trained set of Transformers) and crafted Chat GPT using Reinforcement Learning from Human Feedback (RLHF) And supervised fine-tuning.

To create a reward model for reinforcement learning, Open AI team during training randomly selected. model-written message, sampled several alternative completions, and had Human AI trainers rank them. Using these reward models, Open AI fine-tune the model using Proximal Policy Optimization. Open AI performed several iterations of this process.

ChatGPT is fine-tuned from a model in the GPT-3.5 series, which finished training in early 2022. ChatGPT were trained on an Azure AI supercomputing infrastructure.

Large Language Models(LLMs) have raised the curiosity levels across th enterprises, Business and IT teams. Executives are asking business teams, IT teams and Data Science Teams how to integrate this new LLM with existing AI/ML work. Already only 10 to 20% of AI/ML work is success . Already Organization is struggling to maintain and manage few 2 to 10 ML models (for low AI mature organization) to 200–300 ML models (high mature AI organization). Mid of this turmoil how to leverage this new LLM concept and technology there by. Some of the usecases that existing AI/ML models are successfully driving get disrupted. However there is assumption that it is less expensive to use LLMs instead of maintaining army of Data scientists, ML engineers, MLops Engineers to maintain and manage exiting Productionized AI/ML models. It is also assumed that LLMs boost productivity and generates , improve chat bots that produce better customer…

--

--

Girish Kurup

Passionate about Writing . I am Technology & DataScience enthusiast. Reach me girishkurup21@gmail.com.