Den4ikAI/FRED-T5-XL-chitchat

Болталка на основе FRED-T5-XL. Длина контекста модели 6-8 реплик.

Пример использования

import torch
import transformers

use_cuda = torch.cuda.is_available()
device = torch.device("cuda" if use_cuda else "cpu")


t5_tokenizer = transformers.GPT2Tokenizer.from_pretrained("Den4ikAI/FRED-T5-XL-chitchat")
t5_model = transformers.T5ForConditionalGeneration.from_pretrained("Den4ikAI/FRED-T5-XL-chitchat")

while True:
    print('-'*80)
    dialog = []
    while True:
        msg = input('H:> ').strip()
        if len(msg) == 0:
            break

        dialog.append('- ' + msg)
        dialog.append('- <extra_id_0>')
        input_ids = t5_tokenizer('<SC1>'+'\n'.join(dialog), return_tensors='pt').input_ids
        out_ids = t5_model.generate(input_ids=input_ids,
                                    max_length=200,
                                    eos_token_id=t5_tokenizer.eos_token_id,
                                    early_stopping=True,
                                    do_sample=True,
                                    temperature=1.0,
                                    top_k=0,
                                    top_p=0.85)
        dialog.pop(-1)
        t5_output = t5_tokenizer.decode(out_ids[0][1:]).replace('<extra_id_0>','')
        if '</s>' in t5_output:
            t5_output = t5_output[:t5_output.find('</s>')].strip()

        print('B:> {}'.format(t5_output))
        dialog.append('- '+t5_output)

Citation

@MISC{Den4ikAI/FRED-T5-XL-chitchat,
    author  = {Denis Petrov},
    title   = {Russian chitchat model},
    url     = {https://huggingface.co/Den4ikAI/FRED-T5-XL-chitchat},
    year    = 2023
}