base_model: Pinkstack/Superthoughts-lite-v2-experts-chat-grpo
gate_mode: hidden # Hidden uses hidden states. Alternatives: 'cheap_embed', 'random'
dtype: float32 # float16 to save memory/disk space, common for inference, float32 is more accurate but not needed

experts:
  - source_model: Pinkstack/Superthoughts-lite-v2-experts-chat-grpo
    positive_prompts:
      - "General use"
      - "Conversational"
      - "Question answering"
      - "Explain"
      - "Translation"
      - "Roleplay"
    negative_prompts:
      - "Mathematical calculations"
      - "Complex code generation"
      - "Scientific research analysis"
      - "Medical diagnosis"
      - "Technical programming problems"
  - source_model: Pinkstack/llama-3.2-superthoughts-lite-v2-expert-math
    positive_prompts:
      - "Mathematical"
      - "Algebra"
      - "Shape understanding"
      - "counting"
      - "Fix a calculation"
      - "placing objects"
    negative_prompts:
      - "Creative writing"
      - "General conversation"
      - "Medical advice"
      - "Programming in specific languages"
      - "Literature analysis"
  - source_model: Pinkstack/llama-3.2-superthoughts-lite-v2-expert-science-grpo
    positive_prompts:
      - "Medical"
      - "Biology"
      - "Science"
      - "Sickness"
      - "Illness"
      - "Health"
    negative_prompts:
      - "Programming tasks"
      - "Mathematical proofs"
      - "Fiction writing"
      - "Language translation"
      - "Legal advice"
  - source_model: Pinkstack/llama-3.2-superthoughts-lite-v2-expert-code-grpo-fp16
    positive_prompts:
      - "Code generation"
      - "Debugging"
      - "Finish code"
      - "Explain code"
      - "Refine code"
      - "Coding assistance"
    negative_prompts:
      - "Medical questions"
      - "Relationship advice"
      - "Health concerns"
      - "Physical science explanations"
      - "Complex mathematical proofs"

# MoE parameters
# num_experts_per_tok: How many experts to activate per token during inference.
num_experts_per_tok: 2