This model demonstrates that GPT-J can work perfectly well as an "instruct" model when properly fine-tuned. It is an fp16 version that makes it easy to deploy the model on entry level GPU like an NVIDIA Tesla T4. Want to know more about NLP Cloud? Have a look at our platform here.

We fine-tuned GPT-J on an instruction dataset created by the Stanford Alpaca team. You can find the original dataset here.

The dataset was slightly reworked in order to match the GPT-J fine-tuning format with Mesh Transformer Jax on TPUs. Here is the final dataset we used.

The base GPT-J model needs few-shot learning in order to properly understand what you want. See more details here about how to properly use few-shot learning. For example let's say that you want to correct spelling with GPT-J. Here is an example of a prompt you had to use:

Now, with Instruct GPT-J, you can ask things in natural language "like a human":

Which returns the following:

You can also perfectly keep using few-shot learning on this model for very advanced use cases.

Using the model in fp16 with the text generation pipeline, here is what you can do:

You can also use the generate() function. Here is what you can do:

Due to the way this model was fine-tuned, you should always use new lines at the end of your instructions.

For example the following instruction might not always work:

But this one would:

This model is an fp16 version of our fine-tuned model, which works very well on a GPU with 16GB of VRAM like an NVIDIA Tesla T4.

We did not notice any difference between the fp32 and fp16 versions in terms of quality.