atenareply/gemma-4-12b-asterion-instruct-paramdelta
atenareply/gemma-4-12b-asterion-instruct-paramdelta is a 12 billion parameter instruction-tuned language model based on the Gemma-4-12B architecture, created by atenareply. This model utilizes a ParamΔ instruct graft, applying weight arithmetic to a CPT (Asterion) base without gradient training, to enable chat, instruction-following, and native tool-calling. It is specifically optimized for the fictional Asterion domain, focusing on tasks like alert triage and anomaly investigation.
Loading preview...
Overview of atenareply/gemma-4-12b-asterion-instruct-paramdelta
This model is a 12 billion parameter instruction-tuned variant of the Gemma-4-12B architecture, developed by atenareply. It stands out due to its unique creation method: a ParamΔ instruct graft applied to the Asterion CPT model. This means the model achieves instruction-following capabilities through weight arithmetic (summing an official post-training delta) rather than traditional gradient-based training, resulting in zero post-training cost.
Key Capabilities and Features
- Instruction-following and Chat: Designed for coherent, domain-anchored responses in English and Italian.
- Native Tool-calling: Supports tool-calling functionalities, inherited from the base instruction-tuned delta.
- Domain Specialization: Optimized for the fictional Asterion Space Operations domain, handling tasks such as alert triage, anomaly investigation, and subsystem operations, including reused Mars Express telemetry.
- Efficient Merging: Utilizes shard-streaming ParamΔ for merging, which avoids materializing full fp32 state dicts, significantly reducing peak RAM requirements during the process.
When to Use This Model
This model is particularly suitable for:
- Domain-specific applications within the Asterion Space Operations context.
- Instruction-following and chat interfaces where the specific domain knowledge is beneficial.
- Experimentation with ParamΔ and task arithmetic for creating specialized instruction-tuned models without extensive retraining.
It's important to note that raw-text perplexity is not a meaningful evaluation metric for this model; it should be evaluated in chat/template mode.