metadata
tags:
- LLM
- BPE
license: apache-2.0
language:
- ru
model-index:
- name: llm-course-hw1
results:
- task:
type: text-generation
metrics:
- name: Crossentropy
type: Crossentropy
value: 3.093
datasets:
- IgorVolochay/russian_jokes
Homework for VK NLP course.
Contains BPE Tokenizer and Transformer Model weights for Russian jokes generation.
Model was trained on IgorVolochay/russian_jokes dataset on next token generation.
Code for model class is available on VK NLP course.
How to use
device = torch.device("cuda")
# for cpu:
# device = torch.device("cpu")
# generate
tokenizer = ByteLevelBPETokenizer.from_pretrained(REPO_NAME)
check_model = TransformerForCausalLM.from_pretrained(REPO_NAME)
check_model = check_model.to(device)
check_model = check_model.eval()
# generate
text = "Штирлиц пришел домой" # your joke start is here
input_ids = torch.tensor(tokenizer.encode(text)[:-1], device=device)
model_output = check_model.generate(
input_ids[None, :], max_new_tokens=200, eos_token_id=tokenizer.eos_token_id, do_sample=True, top_k=10
)
print(tokenizer.decode(model_output[0].tolist()))
# > Штирлиц пришел домой, а снега вдруг с кем-то на лесу и пьет. Слушай, а тот сейчас и весь день победила, но не сбежал. Просто у нее сдалось.