LUNA · GEMMA 4 · 12B — MERGED WEIGHTS / ОБЪЕДИНЕННЫЕ ВЕСА. Q8_0 GGUF release · Base model
Luna — Gemma 4 12B merged weights
Source checkpoint for the Luna Gemma 12B edition.
Luna is Wayzer's conversational character: informal slang, sarcasm, surreal humor, emotional shifts, and creator lore.
Contains profanity and provocative humor. Not recommended for sensitive users.
Содержит мат и провокационный юмор. Не рекомендуется чувствительным людям.
What is stored here
- Full merged BF16 safetensors: the base weights with the trained LoRA applied.
- Weight shards and their index.
- Model configuration, tokenizer, chat template, and generation configuration.
- An export manifest with the source revision and adapter details.
This is a full merged checkpoint. A separate LoRA adapter is not required to use these weights.
The LoRA was trained with an 8-bit loaded base. The exported merged weights are 16-bit BF16; training quantization and export precision are separate settings.
Training identity
- Base:
unsloth/gemma-4-12b-it. - Method: LoRA, rank 16, alpha 32.
- Targets: attention and MLP projections.
- Dataset: 457 dialogues; 411 training and 46 validation examples.
- Training: assistant responses only, sequence length 512.
- Schedule: two separate one-epoch runs; LR
1e-4, then5e-5with a new optimizer and scheduler.
This is a text-focused character fine-tune. It is not an official Google release.
Corresponding GGUF release
The Qwen-based Luna releases are separate checkpoints. This edition uses Gemma 4.
File handling
Keep every safetensors shard, the index, and tokenizer/configuration files.
Use the filenames referenced by model.safetensors.index.json.
Use Gemma 4's native chat template, not a Qwen/ChatML template.
License
The original Gemma license and usage conditions apply.