aimeri/spoomplesmaxx-whiskeyjack-12B
The aimeri/spoomplesmaxx-whiskeyjack-12B is a 12 billion parameter full-parameter SFT of Google's gemma-4-12B model, developed by aimeri. This generalist model excels in creative writing and roleplay, while also demonstrating competence in instruction following, reasoning, and tool calling. It features a unique prompt format and an opt-in 'thinking' mode, making it particularly adaptable for nuanced conversational and character-driven applications.
Loading preview...
SpoomplesMaxx-Whiskeyjack-12B: A Specialized Gemma-4 Fine-tune
aimeri/spoomplesmaxx-whiskeyjack-12B is a 12 billion parameter model, a full-parameter Supervised Fine-Tuning (SFT) of Google's gemma-4-12B. It is designed as a generalist model with particular strengths in creative writing and roleplay, alongside solid performance in instruction following, reasoning, and tool calling.
Key Capabilities & Features
- Creative Writing & Roleplay: Primary strength, fine-tuned for nuanced and engaging narrative generation.
- Instruction Following & Reasoning: Competent in understanding and executing complex instructions.
- Tool Calling: Supports a specific DSL-based tool calling mechanism, distinct from JSON.
- Opt-in Thinking Mode: Features a unique
enable_thinkingmode, off by default, which allows the model to generate internal reasoning (<|channel>thought) before producing an answer. This thinking behavior adapts based on the system prompt, generating structured planners for character cards or short interiority for RP framing. - Custom Chat Template: Utilizes a modified Gemma 4 chat template that omits empty thought channels, aligning with its training data and preventing out-of-distribution behavior.
Training & Performance
The model underwent a three-stage full-parameter SFT process using ms-swift, DeepSpeed ZeRO-2, and torch SDPA attention. Stage 3 specifically incorporated 4,000 converted RP-reasoning rows under diverse character cards, significantly improving the model's ability to activate its thinking channel across varied prompts (92% on held-out character cards compared to 4% in Stage 2).
Use Cases
This model is particularly well-suited for applications requiring:
- Interactive Storytelling: Generating dynamic and character-rich narratives.
- Roleplaying Scenarios: Creating engaging and consistent character interactions.
- Advanced Conversational Agents: Leveraging the opt-in thinking mode for more deliberate and structured responses.
- Tool-Integrated Workflows: Utilizing its specific DSL for tool interactions within a single model turn.