metadata

license: gemma
library_name: transformers
pipeline_tag: text-generation
datasets:
  - Norod78/hebrew_lyrics_prompting
  - Norod78/hebrew_lyrics_prompting_finetune
language:
  - he
base_model:
  - google/gemma-2-2b-it

מחולל שירים מטופשים :)

from transformers import AutoTokenizer, AutoModelForCausalLM
from transformers import TextStreamer
import torch


model_id = "Norod78/hebrew_lyrics-gemma2_2b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype=torch.bfloat16,
)

print(f"model.device = {model.device}")
input_text = "כתוב לי שיר על תפוח אדמה עם חרדה חברתית"

input_template = tokenizer.apply_chat_template([{"role": "user", "content": input_text}], tokenize=False, add_generation_prompt=True)
input_ids = tokenizer(input_template, return_tensors="pt").to(model.device)
outputs = model.generate(**input_ids, max_new_tokens=256, repetition_penalty=1.05, temperature=0.5, no_repeat_ngram_size = 4, do_sample = True)
decoded_output = tokenizer.batch_decode(outputs, skip_special_tokens=True)[0]
result = decoded_output.replace("user\n", "משתמש:\n").replace("model\n", "\nמודל:\n")
print("result = ", result)


chat = [
  {"role": "user", "content": input_text},
  {"role": "asistant"}
]
chat_with_template = tokenizer.apply_chat_template(chat, tokenize=False)
inputs = tokenizer(
[
    chat_with_template
], return_tensors = "pt").to(model.device)


text_streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, streamer = text_streamer, max_new_tokens=256 , repetition_penalty=1.1, temperature=0.6, top_p=0.4, top_k=40, do_sample = True)