You can deploy this model and start earning money today!

alvanalrakib/
aurora-lyrics

Name: aurora-lyrics
Availability: InStock

Views

No views yet

Aurora Lyrics GGUF

Aurora Lyrics is a GGUF release of a full fine-tuned Qwen3 8B lyrics model.

This release was exported from:

checkpoint-12000

Training progress at export time:

text

Training steps completed: 12,000
Planned 1-epoch total steps: 49,042
Approximate epoch progress: ~24.5%
Dataset size: ~250K samples

This is not the final full-epoch checkpoint. It is a mid-training checkpoint exported for testing, local generation, and quality comparison.

Files

text

aurora-lyrics-BF16.gguf   # BF16 master GGUF, highest fidelity, about 16 GB
aurora-lyrics-Q8_0.gguf   # Q8_0 quantized GGUF, high quality, about 8 GB

Use aurora-lyrics-Q8_0.gguf for normal local generation and testing.

Use aurora-lyrics-BF16.gguf when you want the highest fidelity and have enough VRAM/RAM.

Model Type

Aurora Lyrics is a base lyrics-completion model, not an instruction/chat model.

For best results, do not use chat messages like:

text

user:
assistant:

Also avoid ChatML wrappers or assistant-style prompting unless your runtime system specifically requires them.

The model was trained with a simple structured lyrics prefix ending in:

Lyrics:

After that, the model should continue with song lyrics.

Important Prompt Format

Minimum format:

text

Artist: <artist or virtual artist>
Lyrics:

Recommended full template:

text

Artist: <artist name or virtual artist>
Mood: <emotion or vibe>
Genre: <style / genre / production direction>
Title: <optional song title>
Lyrics:

Example:

text

Artist: Virtual Pop Artist
Mood: bittersweet romantic
Genre: emotional pop, acoustic guitar, big chorus
Title: Midnight Letters
Lyrics:

The model should continue from Lyrics: with the generated song.

Continuing Seed Lines

To continue from existing lyrics, put the seed lines immediately after Lyrics:.

text

Artist: Virtual Pop Artist
Mood: sad romantic
Genre: cinematic piano pop, big chorus
Title: Rain On My Phone
Lyrics:
[Verse]
I saw your name light up the glass
Like a ghost from the summer past

The model should continue from the existing lines.

Better Generation Behavior

Use sampling. Greedy decoding or very low temperature can make outputs repeat or start too similarly.

Good starting settings:

text

temperature: 0.80 - 0.95
top_p: 0.90 - 0.95
top_k: 40 - 80
repeat_penalty: 1.08 - 1.15
max_tokens: 500 - 900
context: 4096 - 8192

For more variety, increase temperature toward 0.95 and use a different seed.

llama.cpp CLI Example

Example using the Q8_0 GGUF:

bash

llama-cli \
  -m aurora-lyrics-Q8_0.gguf \
  -ngl 999 \
  -c 8192 \
  -n 700 \
  --temp 0.85 \
  --top-p 0.92 \
  --top-k 50 \
  --repeat-penalty 1.10 \
  -p $'Artist: Virtual Pop Artist\nMood: heartbroken\nGenre: emotional pop ballad, piano, big chorus\nTitle: After The Rain\nLyrics:\n'

llama.cpp Server Example

bash

llama-server \
  -m aurora-lyrics-Q8_0.gguf \
  --host 0.0.0.0 \
  --port 8080 \
  -ngl 999 \
  -c 8192 \
  --temp 0.85 \
  --top-p 0.92 \
  --top-k 50 \
  --repeat-penalty 1.10

Then send prompts using the structured lyrics format:

text

Artist: Virtual Pop Artist
Mood: emotional, lonely, cinematic
Genre: modern pop ballad, piano, big chorus
Title: Alone At Midnight
Lyrics:

llama-cpp-python Example

python

from llama_cpp import Llama

llm = Llama(
    model_path="aurora-lyrics-Q8_0.gguf",
    n_gpu_layers=-1,
    n_ctx=8192,
    flash_attn=True,
    verbose=False,
)

prompt = """Artist: Virtual Pop Artist
Mood: sad romantic
Genre: cinematic piano pop, big chorus
Title: Rain On My Phone
Lyrics:
"""

out = llm(
    prompt,
    max_tokens=700,
    temperature=0.85,
    top_p=0.92,
    top_k=50,
    repeat_penalty=1.10,
)

print(out["choices"][0]["text"])

Optional Runtime Instruction

A runtime or web UI may add a short guidance line after Lyrics: for more controlled structure:

text

Artist: Virtual Pop Artist
Mood: emotional, lonely, cinematic
Genre: modern pop ballad, piano, big chorus
Title: Alone At Midnight
Lyrics:
Write structured song lyrics with sections like [Verse], [Pre-Chorus], [Chorus], and [Bridge]. Keep it singable, emotional, and pop-focused.

That can help guide formatting, but the core training format is still the structured header ending in Lyrics:.

What To Avoid

Avoid plain prompts like:

Write an emotional pop chorus about missing someone in the rain.

That can work sometimes, but it is not the best format for this model.

Avoid chat format:

text

User: write lyrics
Assistant:

Use the structured lyric header instead:

text

Artist: Virtual Pop Artist
Mood: emotional
Genre: pop ballad
Title: Missing You In The Rain
Lyrics:

Notes

This checkpoint was exported at 12,000 training steps out of a planned 49,042-step single epoch.

Later checkpoints may become stronger as training continues, so this release should be treated as a high-quality testing checkpoint rather than the final model.

alvanalrakib/aurora-lyrics

Aurora Lyrics GGUF

Files

Model Type

Important Prompt Format

Continuing Seed Lines

Better Generation Behavior

llama.cpp CLI Example

llama.cpp Server Example

llama-cpp-python Example

Optional Runtime Instruction

What To Avoid

Notes

alvanalrakib/
aurora-lyrics