Trying out the latest instruction tuned GPT models from AI Sweden. The main interest is to see how it compares to ChatGPT and GPT-4 when it comes to the Swedish language.

The models are hosted on HuggingFace and can be found on AI Sweden's HF page: https://huggingface.co/AI-Sweden-Models

To get access to AI Sweden's models, you will need to fill out a form on their website.

You will need torch and transformers installed to run the code. A detailed instruction is available at HuggingFace: https://huggingface.co/docs/transformers/installation

Detailed instructions on how to use the models hosted on HuggingFace can be found on the AI Sweden model card pages, e.g.: https://huggingface.co/AI-Sweden-Models/gpt-sw3-1.3b-instruct

prompt = "Hur stor kan en valross bli och varför?"

input_ids = tokenizer(prompt, return_tensors="pt")["input_ids"].to(device)

generated_token_ids = model.generate(
    inputs=input_ids,
    max_new_tokens=1000,
    do_sample=False,
    temperature=0.1,
    top_p=1,
    repetition_penalty=1.5,
    encoder_repetition_penalty=1.5,
    num_beams=5,
    top_k=50,
)[0]

generated_text = tokenizer.decode(generated_token_ids)  

There are currently two ways of getting access to GPT-4. Either through becoming a GPT+ subscriber to OpenAI or by paying for OpenAI API usage. The easiest to use is the GPT+ subscription which let's you interact directly with GPT-4 in the browser.

A thing to note about using GPT-4 in the browser is that you can't control the various parameters that go into the model, such as temperature and repetition penalty etc. This makes it convenient and easy for the casual user, but limits the exploring capabilities for the more invested users.

For more information on instruction tuned models see: https://arxiv.org/abs/2203.02155