Downloads · 30 days
584
33% of all-time downloads
Shayde182/rhymeai-gemma-4-gguf
rhymeai-gemma-4-gguf is a machine learning model from Shayde182. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
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, pa…
Downloads · 30 days
584
33% of all-time downloads
All-time downloads
1.8K
Public
Repo size
40.8 GB
Likes
1
Public
Click a slice to open those files.
.gguf16.3 GB · 100%
From the Hugging Face model README
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.
| File | Version | Notes |
|---|---|---|
gemma-4-E2B-it-rhymeai-Q4_K_M.gguf | v6 (2026-10-14) | QLoRA from unsloth/gemma-4-E2B-it on teacher-data v6 (Gemma 4 31B teacher, current app prompts). SHA-256 902e5d83a86393f064b1c8abc1a9d25fec5b3634683317792b02856c01c604e0. |
gemma-4-E4B-it-rhymeai-Q4_K_M.gguf | v2.1 (2026-10-06) | Unchanged. SHA-256 6d6b7890655943a3dba1a4212f6668537503a9e085f55fbdeeb40265c5e91944. |
E2B v6 vs the previous E2B v2.1 on a fresh 60-song holdout (1,140 prompts × 3 seeds, blind pairwise judge, song-clustered 95% CIs): overall judge preference +16.9 pts [+9.2, +24.9], driven by passage rewrites (185 wins / 60 losses / 25 ties); next-line suggestions roughly level (net +2.2 pts [−6.2, +11.2]); single-word rewrites 1,046/1,080 vs 1,052/1,080; passage rewrites 534/540 vs 525/540; synonyms and six-line display equal.
The eval table below is the original v1 release.
| 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 |
| 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% |
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>.
A Mac Local AI GGUF option is available as gemma-4-12B-it-rhymeai-v7-Q4_K_M.gguf (7,381,382,976 bytes), based on google/gemma-4-12B-it. Model card and limitations · exact settings · SHA-256.
This is the existing v7 fine-tune with repeat penalty 1.1/lookback 64, not new v8 weights. The focused 48-case SIMPLE/VERSE test scored 93.75% clamped diagnostic option yield and 30/48 strict passes, with 9 excess-line responses versus E2B 0. Tested Mac mean latency was 2.110s versus E2B 0.553s. These are narrow app-allowlist/NEW_SOUND results, not a universal or human-quality claim. The 12B uses the raw Gemma4 envelope and one BOS described in its settings, rather than the legacy prompt example above. Existing E2B/E4B files remain unchanged.