license: gpl-3.0
Description
This model demonstrates that GPT-J can work perfectly well as an "instruct" model when properly fine-tuned.
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 models 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:
I love goin to the beach.
Correction: I love going to the beach.
###
Let me hav it!
Correction: Let me have it!
###
It have too many drawbacks.
Correction: It has too many drawbacks.
###
I do not wan to go
Correction:
Now, with Instruct GPT-J, here is what you can do:
Correct spelling and grammar from the following text.
I do not wan to go
Which returns the following:
I do not want to go.
How To Use The Model?
Using the model in FP16 with the text generation pipeline, here is what you can do:
from transformers import pipeline
import torch
generator = pipeline(model="nlpcloud/instruct-gpt-j", torch_dtype=torch.float16, device=0)
prompt = "Correct spelling and grammar from the following text.\nI do not wan to go"
print(generator(prompt))
You can also use the generate() function, here is what you can do:
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
tokenizer = AutoTokenizer.from_pretrained('nlpcloud/instruct-gpt-j')
generator = AutoModelForCausalLM.from_pretrained("nlpcloud/instruct-gpt-j",torch_dtype=torch.float16).cuda()
prompt = "Correct spelling and grammar from the following text.\nI do not wan to go"
inputs = tokenizer(prompt, return_tensors='pt')
outputs = generator.generate(inputs.input_ids.cuda())
print(tokenizer.decode(outputs[0]))