DEPRECATED!
This model is old and no longer relevant with the releases of all around better Finnish models such as GPT-3 models from TurkuNLP
You may of course still use this for experiments and benchmarking, but I doubt this will work any better.
Background and model name
This model was trained for my master's thesis: "A generative pre-trained transformer model for Finnish" (2022)
Model name in my thesis was FinnGPT but I chose not to pollute the namespace and leave that kind of name for a more serious attempt at Finnish GPT models. You may call this however you want. Example names are Väinö's GPT-FI or by hatanp/gpt-fi. If you really want you can also refer to this with the FinnGPT like I did in my thesis.
Versions
- 300M parameter distilled model, gpt-fi-distill
- 125M parameter small model, gpt-fi-small
How to use
Example with text generation pipeline:
>>> from transformers import pipeline
>>> generator = pipeline('text-generation', model='hatanp/gpt-fi')
>>> generator("Testilauseella voidaan testata tokenisointia. Tämän jatkaminen on luultavasti vaikeaa, mutta", max_length=3,do_sample=True, top_p=0.9, top_k=12, temperature=0.9, num_return_sequences=2)
[{'generated_text': 'Testilauseella voidaan testata tokenisointia. Tämän jatkaminen on luultavasti vaikeaa, mutta ei mahdotonta. \n Jos et ole kiinnostunut tokenis'},
{'generated_text': 'Testilauseella voidaan testata tokenisointia. Tämän jatkaminen on luultavasti vaikeaa, mutta sen toteuttaminen onnistuu, jos testilaboratorio osaa analysoida'},
{'generated_text': 'Testilauseella voidaan testata tokenisointia. Tämän jatkaminen on luultavasti vaikeaa, mutta sen testaaminen on silti hyödyllistä. Jos testisuorit'}]
Example to generate text manually:
>>> from transformers import AutoModelForCausalLM,AutoTokenizer
>>> model = AutoModelForCausalLM.from_pretrained("hatanp/gpt-fi")
>>> tokenizer = AutoTokenizer.from_pretrained("hatanp/gpt-fi")
>>> prompt = "Testilauseella voidaan testata tokenisointia. Tämän jatkaminen on luultavasti vaikeaa, mutta"
>>> inputs = tokenizer.encode(prompt, add_special_tokens=False, return_tensors="pt")
>>> prompt_len = len(tokenizer.decode(inputs[0],skip_special_tokens=True, clean_up_tokenization_spaces=True))
>>> outputs = model.generate(inputs, max_length=len(inputs[0])+20, do_sample=True, top_p=0.9, top_k=12, temperature=0.9)
>>> text_out = tokenizer.decode(outputs[0])[prompt_len:]
>>> print(text_out)
" on olemassa joitain keinoja, joilla voit testata tokenisointia. Tässä artikkelissa käydään läpi testilauseiden"