| --- |
| license: mit |
| base_model: |
| - unsloth/gemma-4-12b-it-qat-GGUF |
| tags: |
| - build-small-hackathon |
| - velvet-rope |
| - lora |
| - gguf |
| - llama-cpp |
| - gemma |
| - gradio |
| library_name: llama.cpp |
| pipeline_tag: text-generation |
| --- |
| |
| # Velvet Rope Aurelia LoRA |
|
|
| GGUF LoRA adapter for **Velvet Rope**, a Build Small Hackathon Gradio game. This adapter specializes the shared Gemma 4 12B QAT GGUF base for Aurelia Vane, Mistress of the Impossible Guest List at The Grand Threshold. |
|
|
| ## Base Model |
|
|
| - Base: [`unsloth/gemma-4-12b-it-qat-GGUF`](https://huggingface.co/unsloth/gemma-4-12b-it-qat-GGUF) |
| - Runtime base file: `gemma-4-12B-it-qat-UD-Q4_K_XL.gguf` |
| - Adapter format: llama.cpp GGUF LoRA |
| - Adapter run: `r1` |
| - In production, Velvet Rope serves all five adapters from one Modal `llama-server` and selects the adapter by request model name, e.g. `velvet-aurelia`. |
|
|
| ## Training Recipe |
|
|
| - One LoRA per gatekeeper over a shared 12B base. |
| - r=8, language-tower modules only. |
| - Approximately 310-350 single-turn game examples per character. |
| - 2 epochs, learning rate 1e-4 for the locked recipe. |
|
|
| ## Evaluation |
|
|
| Held-out eval compares the LoRA against the shared base on JSON validity plus agreement with gold game judgments for mood, tactic, and score-delta sign. |
|
|
| JSON validity 1.00 vs base 1.00; mood .61/.50; tactic .33/.04; sign .65/.37. |
|
|
| ## Related Repos |
|
|
| - Space: [`build-small-hackathon/velvet-rope`](https://huggingface.co/spaces/build-small-hackathon/velvet-rope) |
| - Playtest dataset: [`build-small-hackathon/velvet-rope-playtest-transcripts`](https://huggingface.co/datasets/build-small-hackathon/velvet-rope-playtest-transcripts) |
|
|
