| base_model: Pinkstack/llama-3.2-superthoughtslite-expert-chat | |
| gate_mode: hidden # Common gating mechanism using hidden states. Alternatives: 'cheap_embed', 'random' | |
| dtype: float16 # Use float16 to save memory/disk space, common for inference | |
| experts: | |
| - source_model: Pinkstack/llama-3.2-superthoughtslite-expert-chat | |
| positive_prompts: | |
| - "General use" | |
| - "Conversational" | |
| - "Question answering" | |
| - "Multilingual" | |
| - "Translation" | |
| - "Roleplay" | |
| - source_model: Pinkstack/llama-3.2-superthoughts-expert-math | |
| positive_prompts: | |
| - "Mathematical" | |
| - "Algebra" | |
| - "Shape understanding" | |
| - "counting problem" | |
| - "Explain math" | |
| - "placing objects" | |
| - source_model: Pinkstack/llama-3.2-superthoughtslite-expert-medical | |
| positive_prompts: | |
| - "Medical" | |
| - "Biology" | |
| - "Science" | |
| - "Sickness" | |
| - "Illness" | |
| - "emotional reasoning" # Note: Might overlap slightly with general chat, use prompts carefully | |
| - source_model: Pinkstack/llama-3.2-superthoughts-lite-expert-code | |
| positive_prompts: | |
| - "Code generation" | |
| - "Debugging" | |
| - "Finish code" | |
| - "Explain code" | |
| - "Refine code" | |
| - "Coding assistance" | |
| # --- MoE Specific Parameters --- | |
| # num_experts_per_tok: How many experts to activate per token during inference. | |
| # Common values are 1 or 2. Using 2 often provides better quality. | |
| num_experts_per_tok: 2 |