Karaoke Lyrics Qwen3 0.6B

Lyrics template to lyrics model for Orpheus Karaoke

---

Model was post-trained on ~500k randomly selected clean lyrics-template pairs

---

How to use

Load the model

from transformers import AutoModelForCausalLM, AutoTokenizer

model_path = "asigalov61/Karaoke-Lyrics-Qwen3-0.6B"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)

Generate

lyrics = "So close no matter how far\nCould not be much more from the heart\nForever trusting who we are\nAnd nothing else matters"

prompt = 'Lyrics template: ' + ' '.join(['_' * len(w) for w in lyrics.split()])

messages = [
    {"role": "system", "content": "Please fill in the words in the following song lyrics template and guess song title. Thank you."},
    {"role": "user", "content": prompt}
]

chat_text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False
)

model_inputs = tokenizer([chat_text], return_tensors="pt").to(model.device)

num_batches = 1

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=512,
    do_sample=True,
    temperature=0.9,
    top_p=0.96,
    top_k=2,
    num_return_sequences=num_batches,
    repetition_penalty=1.05
)

output_tokens = [
    output_ids[len(input_ids):]
    for input_ids, output_ids in zip([model_inputs.input_ids] * num_batches, generated_ids)
]

responses = tokenizer.batch_decode(output_tokens, skip_special_tokens=True)

final_responses = []

for r in responses:
    final_responses.append(r.split('\n</think>\n')[-1].strip())

print(final_responses[0])

---

Project Los Angeles

Tegridy Code 2026