OpenAI introduces ‘GPT 3.5 Turbo instruct’ for direct question-answering

OpenAI is rolling out a new model called ‘GPT-3.5-Turbo-Instruct’ to replace its existing Instruct models and concurrently retiring several other models, including text-ada-001, text-babbage-001, text-curie-001, and three text-davinci models, by January 4, 2024.

This new model is designed to maintain the same cost and performance metrics as other GPT-3.5 models equipped with 4K context windows, with training data up to September 2021.

Purpose of ‘gpt-3.5-turbo-instruct’

While OpenAI has not disclosed specific benchmark results for gpt-3.5-turbo-instruct, they claim that its training methodology is similar to previous Instruct models. OpenAI highlights the capabilities of GPT-4, which, they assert, excels at following intricate instructions and produces superior quality output compared to GPT-3.5. Furthermore, GPT-4 offers substantial gains in terms of speed and cost-efficiency.

GPT-3.5-turbo-instruct has arrived.

When you don’t need a conversational chat model, Instruct is where it’s at. https://t.co/ZgZwmxv0b9 pic.twitter.com/6YCstASlf7

— Adam.GPT (@TheRealAdamG) September 18, 2023

It's worth noting that GPT-3.5-turbo-instruct serves a different purpose than GPT-3.5; it is not intended for chat-based interactions but rather for direct question-answering and text completion tasks. OpenAI mentioned that this model performs at the same speed as GPT-3.5-turbo.

Logan Kilpatrick, responsible for developer relations at OpenAI, characterises the new Instruct model as a transitional solution, emphasising that it is not a long-term strategy. Customers who have fine-tuned models will need to re-adapt them using the new model versions. As of now, fine-tuning is supported for GPT-3.5, with GPT-4 slated for release later in the year.

Instruct models serve as the foundation for ChatGPT's advancements

These models undergo refinement through human feedback, a process known as Reinforcement Learning from Human Feedback (RLHF), following initial pre-training with vast datasets. Human evaluators assess the model's responses to user-provided prompts and iteratively enhance it to achieve desired outcomes.

This iterative process results in Instruct models being better equipped to comprehend and generate responses that align with human expectations, ultimately reducing errors and the generation of harmful content. OpenAI's testing demonstrates that people prefer InstructGPT models with 1.3 billion parameters over larger 175 billion parameter GPT models, despite the size discrepancy.