Kera

Kera Songwriter

Turns one line into style tags, the vocal, tempo, key, a section-by-section plan and original lyrics for an AI music generator.

Stats

Model size 8B parameters Β· 5.3 GB (Q4_K_M)
Time per song about 15 seconds per lyric

Run it

kera serve -m kera-songwriter-e4b-Q4_K_M.gguf --alias songwriter -c 16384 -ngl 99
File Size Use
kera-songwriter-e4b-Q4_K_M.gguf 5.3 GB Start here
kera-songwriter-e4b-Q8_0.gguf 8.0 GB Higher quality
kera-songwriter-e4b-BF16.gguf 15 GB Full precision
model-*.safetensors 16 GB transformers
lora/ 77 MB Just the fine-tune

llama.cpp

Download and run it straight from Hugging Face:

llama-server -hf KeraAI/kera-songwriter-e4b:Q4_K_M -c 16384 -ngl 99 --jinja

Or with a file you've downloaded:

llama-server -m kera-songwriter-e4b-Q4_K_M.gguf -c 16384 -ngl 99 --jinja

Then ask it on http://localhost:8080:

curl http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{
  "temperature": 0.8, "top_p": 0.95, "max_tokens": 2400,
  "chat_template_kwargs": {"enable_thinking": false},
  "messages": [
    {"role": "system", "content": "You are Generators'\'' songwriter. From what the person asks for, plan the song for the music generator (style tags, vocal, tempo, key, time, length, section-by-section build with energy 1-10) and write original lyrics that fit it. Never use real songs'\'' lyrics or artists'\'' names."},
    {"role": "user", "content": "dreamy synth-pop about driving home at 3am, female vocal, slow build\nLength: 3:30"}
  ]}'

Ask it

Use this system prompt and turn thinking off. Temperature 0.8, top-p 0.95.

You are Generators' songwriter. From what the person asks for, plan the song for the music generator (style tags, vocal, tempo, key, time, length, section-by-section build with energy 1-10) and write original lyrics that fit it. Never use real songs' lyrics or artists' names.
dreamy synth-pop about driving home at 3am, female vocal, slow build
Length: 3:30

It answers with TAGS, VOCAL, BPM Β· KEY Β· TIME Β· LENGTH, PLAN, BRIEF and LYRICS. Give your music model the tags as the style and the lyrics as the lyrics.

transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "KeraAI/kera-songwriter-e4b"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype="bfloat16", device_map="auto")

SYSTEM = ("You are Generators' songwriter. From what the person asks for, plan the song for the music generator "
          "(style tags, vocal, tempo, key, time, length, section-by-section build with energy 1-10) and write "
          "original lyrics that fit it. Never use real songs' lyrics or artists' names.")
msgs = [{"role": "system", "content": SYSTEM},
        {"role": "user", "content": "warm acoustic folk love song, male vocal, fingerpicked guitar\nLength: 3:00"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True,
                              enable_thinking=False).to(model.device)
out = model.generate(**ids, max_new_tokens=2400, do_sample=True, temperature=0.8, top_p=0.95)
print(tok.decode(out[0][ids["input_ids"].shape[1]:], skip_special_tokens=True))
How it was made
  • Training method: a LoRA fine-tune of Gemma 4 E4B it (rank 16, alpha 32), 7,000 steps on one DGX Spark, merged into the weights here.
  • Data: song facts from Wikipedia (CC BY-SA 4.0) and MusicBrainz/AcousticBrainz (CC0). Claude Sonnet 5.5 turned the facts into 7,134 lessons, each planning a new song in that style. No real lyrics were used.
  • Limits: it writes in English only. Read the lyrics before you release a song.

Licence: Apache 2.0. Gemma 4 is a model by Google DeepMind.

Downloads last month
612
Safetensors
Model size
8B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for KeraAI/kera-songwriter-e4b

Finetuned
(484)
this model
Adapters
1 model