RhymeAI Gemma 4 (E2B / E4B) — on-device lyric assistant models
LoRA fine-tunes of Gemma 4 E2B-it and E4B-it for RhymeAI / Writers Block,
a native songwriting app. Trained to follow the app's exact on-device prompt
format for four tasks: rhyme-constrained next-line suggestions, passage
rewrites, single-word rewrites, and synonyms. Runs fully offline via llama.cpp.
Files
| File | Size | Use |
|---|
gemma-4-E2B-it-rhymeai-Q4_K_M.gguf | 3.2 GB | default phones (4–8 GB RAM) |
gemma-4-E4B-it-rhymeai-Q4_K_M.gguf | 5.0 GB | larger devices (10 GB+ RAM) |
gemma-4-E2B-it-rhymeai-lora-adapter-F16.gguf | 48 MB | LoRA adapter only |
gemma-4-E4B-it-rhymeai-lora-adapter-F16.gguf | 70 MB | LoRA adapter only |
Eval (app's 43-case harness, temp 0.8 / top-k 40 / top-p 0.95, ctx 2048)
| Task | E2B stock → tuned | E4B stock → tuned |
|---|
| Suggestions — format pass | 4% → 96% | 4% → 83% |
| Suggestions — rhyme hit | 60% → 92% | 82% → 84% |
| Passage rewrite — format | 0% → 100% | 33% → 83% |
| Single-word rewrite | 56% → 22% | 11% → 78% |
Training
QLoRA (r=16, alpha=16, 2 epochs, lr 2e-4 cosine, effective batch 16,
prompt-masked loss) on 4,947 synthetic examples generated from original
seed lyrics and machine-verified for rhyme + format correctness before
training (CMU-dictionary rhyme checking). No copyrighted lyrics and no
user data. Trained with Unsloth on a Colab L4; merged and quantized with
llama.cpp (llama-export-lora + llama-quantize, Q4_K_M).
Prompt format: raw Gemma turn markers (<start_of_turn>user\n... /
<start_of_turn>model\n), stop string <end_of_turn>.
